CompareRows checked – · benchmark links added · Kev and Laya rows · Strands Decider row · hosted-API layer, 3 new rows and Strands Decider rechecked · Microsoft-Decision-1 row added and Decisions API evidence rebuilt from graded rows · Microsoft-Decision-1 cells, Decisions API data controls, B41 and Cloudflare Clef family rows · three hosted rows and Clef base licences rechecked, B42 added · SubQ Snap row added, B43 and B44 added · Jev input wording and Microsoft-Decision-1 lifecycle rechecked · Liquid AI d1-ultra and Mouse-1 noted as launched, not yet evaluated, · Liquid AI d1-ultra, Mouse-1 and Eevo rows added · Drex (Nace.AI) row added · Drex legal pages read · OpenAI Decisions API status rechecked · OpenAI Decisions API caching prices, B45 and B46 added · OpenAI Decisions API price and data controls, Microsoft-Decision-1 price, Clef hosted prices and SubQ Snap price rechecked · hosted rows re-read (Eevo price, Microsoft-Decision-1 rate limits, Clef data terms and OpenRouter hosts, OpenAI input and region wording changed) and Kev and Laya releases noted · Laya release v0.4.21 noted; Microsoft-Decision-1 catalog cells, Drex, SubQ Snap and Kev re-read unchanged · Base Labs row added; Laya release v0.4.28 noted; hosted rows re-read
Jev alternatives (TypeSafe AI): open-source models, local runtimes and hosted APIs, and what each replaces
What can replace Jev (TypeSafe AI): open-weight models, local runtimes, request adapters and hosted decision APIs from 9 other providers (OpenAI, Microsoft, Cloudflare, Subquadratic, Liquid AI, Mouse, Upright, Nace.AI, Baseten (Base Labs)), by layer and licence. "Alternative" can mean five different things: a different request library, a local runtime, a different trained model, another provider's hosted API, or a different kind of model altogether. Pick the layer you need to replace first; a matching request shape does not mean matching answers. Why this title: Google completes "jev alt" to "jev alternatives"; Bing and DuckDuckGo complete it to "jev alternative" (autocomplete, 10 Oct 2026, 05:17 UTC; R12-S11). A suggestion shows that people search for it, not how many.
Key finding
Jev weights: not found on TypeSafe's site, docs, GitHub or Hugging Face (10 Oct). Of 39 listed substitutes: 13 model weights, 11 local runtimes, 4 request adapters, 2 classifier/LLM approaches, 9 hosted APIs from other providers (OpenAI, Microsoft, Cloudflare, Subquadratic, Liquid AI, Mouse, Upright, Nace.AI, Baseten (Base Labs)).
Picked from items named in earlier research and the 8–10 Oct scan, added to the rows already on this page, plus Microsoft-Decision-1 from the 11 Oct scan, the Cloudflare Clef family (two rows, hosted and weights) from the 12 Oct review, SubQ Snap (read 14 Oct), Liquid AI d1-ultra, Mouse-1 and Eevo (read 19 Oct), Drex (read 20 Oct) and Base Labs (Baseten; read 7 Nov); not a survey of decision models. Counts are table rows by layer; a row with two layers counts once, under its primary layer. The pool ran nothing: no model was downloaded or run and no Jev call was made.
Change, : 38 → 39 listed substitutes; hosted APIs from other providers 8 → 9 (Base Labs (Baseten): mammoth-1, elephant-1, joey-1 added as one family row from Baseten's library pages, pricing page, Terms, Privacy Policy and Model APIs docs and the Hugging Face cards; R27-S200 to R27-S220, R27-S42). Its MIT weights are a line inside that row, not counted under model weights. The other layer counts are unchanged. Laya row: latest GitHub release v0.4.28 (7 Nov 2026); contents not read (R27-S35). Rechecked and unchanged, re-dated to 7 Nov 2026: d1-ultra price and weights line, Mouse-1 aliases and price, Eevo terms, Drex prices, SubQ Snap status and price (R27-S29 to R27-S33, R27-S47). Not re-read, dates kept: Decisions API retention and ZDR, Clef data usage, SubQ Snap retention. TypeSafe's models page lists the input of both Jev 1.14 and Jev 1.13 as "Text, images" (first observed 7 Nov 2026); the lines that describe Jev's image input now say so (R27-S05, R27-S06). Earlier values kept as dated.
Change, : Laya row: latest GitHub release v0.4.21 (PyPI 0.4.21, 1 Nov 2026, 21:16 UTC); new release, not yet read (R26-S33, R26-S118). Microsoft-Decision-1: the Foundry catalog text was read again and is unchanged: text-only input, 32,768-token context, the licence text and lifecycle Preview (R26-S41). Rechecked and unchanged, re-dated to 2 Nov 2026: Drex prices, SubQ Snap version and prices, Kev's latest release, and the re-read Decisions API and Decision-1 lines. Earlier values kept as dated. Layer counts unchanged (38 rows).
Change, : the Eevo row now gives both published prices: $0.10 per 1M input tokens in Direct mode and the SystemOne-compatible API, and "$0.15 per decision" in Default mode; eevo-1 is now Eevo-1.2 and eevo-1.1 is no longer served; 5 free decisions; SystemOne state up to 130,000 characters (changed between 19 Oct and 31 Oct 2026; R25-S221, R25-S222). Microsoft-Decision-1: Microsoft's quotas page now lists tokens-per-minute and requests-per-minute limits by tier, 2,160 to 15,000 requests per minute (was 60 to 900; changed between 13 Oct and 31 Oct 2026; Microsoft's page shows no new update date; R25-S244). Clef hosted: Cloudflare's data-usage page (updated 16 Oct 2026) now says "Zero data retention: Workers AI processes Customer Content only for as long as needed to provide the service", with exceptions and no period in days stated, and rewords its no-training sentence (R25-S239); on OpenRouter, Clef-flash is now also served by Makora and DeepInfra (changed between 13 Oct and 31 Oct 2026; R25-S42). OpenAI Decisions API: the guide now says it "evaluates text, images, and JSON" and that "Data residency is available in all supported API regions" (changed between 13 and 31 Oct 2026; R25-S206); a per-request image limit was not found in the guide, API reference or changelog on 30 or 31 Oct 2026. Kev and Laya rows: new releases noted, not yet read (R25-S262, R25-S263). Rechecked and unchanged, re-dated to 31 Oct 2026: d1-ultra (its billing sentence is reworded), Mouse-1, Drex prices, SubQ Snap prices, and the other re-read Decisions API, Decision-1 and Clef lines. Earlier values kept as dated. Layer counts unchanged (38 rows).
Change, : the Drex row adds the 30 Oct pricing docs (drex-max $0.03 per 1M input tokens, was $0.06; new drex-max-v1.1 $0.03; signup balance $5.00; changed between 20 and 30 Oct 2026). Also: the OpenAI Decisions API row's data controls now give OpenAI's current wording (Zero Data Retention, Private Retention with PSP and Safety Retention for eligible customers, "see below for limitations"; changed between 13 Oct and 30 Oct 2026; "Pending confirmation" kept as dated; R24-S203). The SubQ Snap row: SubQ's models page lists SubQ Snap 1.1 at $0.04 per 1M input tokens, output free (changed between 14 Oct and 28 Oct 2026; R24-S212); the 12 Oct launch post still says $0.05 (R24-S214). Prices rechecked and unchanged: Decisions API, Microsoft-Decision-1 and Clef hosted (30 Oct 2026, 05:17–05:19 UTC). The Drex row keeps its 20–21 Oct dates. Layer counts unchanged (38 rows).
Change, : the OpenAI Decisions API row's price now gives the caching prices published since general availability (cached input and cache writes are priced; output is still not charged; OpenAI guide and pricing page, 24 Oct 2026, 05:15 UTC; R23-S32, R23-S45, R23-S46); the 13 Oct beta wording "no output, cache-read or cache-write charges" is kept as dated. Graded comparisons with Jev: B45 (grade A) added to the Decisions API row (5 → 6) and B46 (grade A) to the SubQ Snap row (1 → 2). Layer counts unchanged (38 rows).
Change, : the OpenAI Decisions API row and the lists that named it "public beta" now say it became generally available on 22 Oct 2026 (OpenAI API changelog; guide read 23 Oct 2026, 05:23 UTC); requests may now carry up to 400 questions (200 during the beta); model and price unchanged (R22-S32, R22-S47). Counts unchanged.
Change, : the Drex row adds Nace.AI's website Terms and Privacy Policy (9 Oct), which differ from its Console Service Terms and Data Processing Addendum (19 Oct) on training and retention; the drex-v1.5 price units are shown side by side. Layer counts unchanged (38 rows).
Change, : 37 → 38 listed substitutes; hosted APIs from other providers 7 → 8 (Drex from Nace.AI added from its product page, docs, pricing page, terms and model cards; R20-S100 to R20-S131). The other layer counts are unchanged.
Change, : 34 → 37 listed substitutes; hosted APIs from other providers 4 → 7 (Liquid AI d1-ultra, Mouse-1 and Eevo added from their launch posts, docs, pricing pages and data terms; R19-S100 to R19-S152). The 16 Oct Eevo line and the 18 Oct "launched, not yet evaluated" line under seen, not checked are removed; they are rows now. Tazo Labs added there as a lead. The other layer counts are unchanged.
Change, : statements that Jev takes text only are updated: TypeSafe's Jev takes text by default and, since 14 Oct 2026, images as an experimental opt-in, not production-ready per TypeSafe (models and images pages, 16 Oct 2026, 05:21–05:22 UTC; R17-S05, R17-S06). Microsoft-Decision-1: still Preview, version 1, in Foundry (16 Oct 2026, 05:21 UTC; R17-S33). Eevo added to seen, not checked; not a row. Layer counts unchanged (34 rows).
Change, : 33 → 34 listed substitutes; hosted APIs from other providers 3 → 4 (SubQ Snap added from its docs, Terms and launch post; R16-S217 to R16-S227). SubQ Snap removed from 'seen, not checked'. Decisions API and Microsoft-Decision-1 rows: graded comparison B44 added (five and two graded comparisons).
Change, : graded comparisons with OpenAI's Decisions API 3 → 4 (B42 added, grade B). Microsoft-Decision-1: Foundry lifecycle listed as Preview at 13 Oct 2026, 05:20 UTC (listed Generally available on 12 Oct; cause unknown); rate limits by tier, catalog licence text, Product Terms and Microsoft's chart figures added (R15-S44, R15-S200, R15-S206, R15-S209). Decisions API: refusal behaviour, 128-image limit, regional table and Vercel route added (R15-S212, R15-S214, R15-S217). Clef hosted: platform errors and OpenRouter hosts added; data wording now names the host (R15-S223, R15-S226). Clef weights: base-model licences added (R15-S219 to R15-S221). SubQ Snap, Cygnet and GLiNER2.5-Decide added to seen, not checked. Layer counts unchanged (33 rows).
Change, : 31 → 33 listed substitutes; hosted APIs from other providers 2 → 3 and model weights 12 → 13 (Cloudflare Clef family added as two rows, hosted and weights; R14-S400 to R14-S406). Microsoft-Decision-1: request shape, input types, limits, price and a partial data-terms cell added from Microsoft's pages (R14-S207–S215); "terms not read" replaced. Graded comparison B41 added to the Decision-1 and Decisions API rows; OpenAI's data-controls values added to the Decisions API row (R14-S218).
Change, : 30 → 31 listed substitutes; hosted APIs from other providers 1 → 2 (Microsoft-Decision-1 added; R13-S32). The other layer counts are unchanged.
Short answer
Official weights or a local runtime. TypeSafe-published Jev weights or a local Jev runtime: not found on typesafe.ai (no /models page, HTTP 404), the docs (llms.txt, the models page, the introduction), TypeSafe's GitHub organisation (12 repositories: SDKs, an interface adapter, skills, examples, n8n nodes and two forks) or a Hugging Face author search for TypeSafe accounts (0 models), checked . Documented TypeSafe's docs describe Jev as hosted: "the same weights serve every account". Its new Enterprise page lists the question "Can we run Jev in our own VPC or on-premises?", but the answer is not in the served page and the pool has not read it. (R12-S100–S110, R12-S21, R12-S36, R12-S58)
Can you run Jev in Ollama? Not TypeSafe's Jev on the evidence checked. Ollama's library search for "jev" showed no Jev-named model on its first page (10 Oct 2026, 05:11 UTC), and no TypeSafe weights were found (above). Ollama 0.35 (blog, 29 Sep) runs other Jev-style decision models (nimble from Bespoke Labs; tev1 and tev1:0.8b from Together AI) behind a System One-style endpoint. OllamaMQ v0.5.0 is a queue in front of such servers, not a model. Documented (R12-S107, R12-S135, R12-S136)
Other options. If you need local weights, the independently trained options with inspectable repositories and model cards are Kev, CLM, Laya and Jeff, among others; Kev and CLM also expose a System One-style API. Ollaya (new on 28 Sep) runs several of these models locally behind a System One-compatible API. The pool has run no parity test of its own (nothing run as of 10 Oct 2026); their accuracy figures are the authors' own. Published comparisons with Jev: Opper's Jev–Kev comparison (B11, grade B, operator with a stake). Three comparisons of Jev with OpenAI's Decisions API are graded and differ by task: see its row (B38, komzweb, and B39, mateusfonsek, since 11 Oct 2026; B41, kartikanand73, since 12 Oct 2026; all three grade A); on 24 Oct 2026 there are six (B42, grade B, since 13 Oct; B44, grade A, since 14 Oct; B45, grade A, since 24 Oct). One graded comparison includes Microsoft-Decision-1 (two since 14 Oct 2026, with B44): On synthetic bank-withdrawal routing (B41, grade A), Jev (89.2% macro-F1) and Microsoft-Decision-1 (87.6%) were not separated and both were ahead of OpenAI's Decisions API (71.4%); Decision-1 kept the same route across all 6 option orders more often (97.2–97.6% vs Jev 91.6–91.8%, Decisions API 64.2%). Reported; see its row. Hosted Jev, Laya and Kev are compared on the same rows on Jev vs Laya vs Kev. (This paragraph: rows checked 26 Sep–1 Oct, not rechecked on 10 Oct.)
Another provider's hosted API. Another provider's hosted decision API is a different layer. The table lists 9. OpenAI's Decisions API (generally available since 22 Oct 2026) runs its own model, gpt-6-luna, behind a different request shape, and your data goes to OpenAI (row; R12-S134, R11-S300; status R22-S32, R22-S47). Microsoft-Decision-1 (published 9 Oct 2026) is listed in Microsoft Foundry's catalog and on OpenRouter; it is text-only, and your data goes to Microsoft under Microsoft's terms (row; R14-S208, R14-S209, R14-S212, checked 12 Oct 2026, 05:13 UTC). Cloudflare's Clef family (Clef, Clef-flash, Clef-omni) runs on Workers AI with the same state-plus-questions shape as Jev but a different endpoint, and your data goes to Cloudflare under Cloudflare's terms on Workers AI; on OpenRouter's non-Cloudflare hosts (Clef: DeepInfra, PrimeIntellect; Clef-flash: Makora, DeepInfra, PrimeIntellect) it goes to that host (row; R14-S400–S405, checked 12 Oct 2026, 05:33 UTC; OpenRouter endpoints, checked 31 Oct 2026, 05:19 UTC, R25-S42); its weights are also published (row). SubQ Snap (Subquadratic, public preview) uses the same request shape as Jev's /v1/systemone on its own endpoint, with its own model; your data goes to Subquadratic (row; R16-S217 to R16-S227, read 14 Oct 2026, 16:47–16:48 UTC). Liquid AI d1-ultra (available) uses Jev's request shape on its own endpoint, Mouse-1 (preview) and Eevo (Upright) each have their own decision API plus a /systemone endpoint in Jev's shape, and your data goes to each vendor under its own terms, which differ on training and retention (d1-ultra, Mouse-1, Eevo rows; R19-S100 to R19-S152, read 19 Oct 2026, 05:22–05:25 UTC). Drex (Nace.AI; "Generally available" per its launch post) uses Jev's request shape on its own endpoint; Nace's docs call it "wire-compatible" with Jev (a vendor statement, not tested). Its weights are also published, under a modified Open RAIL-M licence with revenue and funding thresholds for commercial use, and your data goes to Nace.AI under its own terms (row; R20-S100 to R20-S128, read 20 Oct 2026, 05:16–05:23 UTC). Base Labs, Baseten's research lab, offers three decision models on Baseten's API: mammoth-1 ($0.10), elephant-1 ($0.035) and joey-1 ($0.01) per 1M input tokens; they are different models, and your data goes to Baseten under Baseten's terms; Base Labs also publishes their weights under MIT (row; R27-S200 to R27-S212, read 7 Nov 2026, 15:30 UTC). All nine are alternatives, not products built on Jev. See the hosted-API layer. Compare side by side: Decisions API vs Jev vs Microsoft Decision 1.
Choosing between Jev, a general LLM with JSON output and a trained classifier is a different question: see Jev vs an LLM vs a classifier: when to use which. Published accuracy tests, graded by how far they can be checked, are on the benchmarks ledger.
Which layer do you need to replace?
Layers an alternative can replace, from least to most
- Your application codeunchanged in every case
- Request shape (API)Your code keeps the same request; a different backend answers4 entriesJevper, asynq-io/system-one, iamaamir/system-one, TypeSafe adapter
- Runtime (where it runs)Runs on your hardware; quality depends on the model it wraps11 entriesOllaya, OpenDecision, LocalJev, reflex and others
- Model weightsA different trained model; answers differ from Jev13 entriesKev, CLM, Laya, Jeff, Strands Decider, Jiwo 0.8B, TOD, Cloudflare Clef weights and others
- Hosted API from another providerA different provider, model and request shape; your data goes to that provider9 entriesOpenAI Decisions API (generally available since 22 Oct 2026), Microsoft-Decision-1, Cloudflare Clef family, SubQ Snap (public preview), Liquid AI d1-ultra, Mouse-1 (preview), Eevo, Drex (Nace.AI), Base Labs (Baseten). Data terms of Jev's own routes: data privacy
- Whole approachA different kind of model or workflow2 entriesGLiClass; an LLM with structured output
| You need | Choose | Do not choose it if |
|---|---|---|
| Keep your request code, change the provider | An API interface adapter | You need data to stay local: adapters forward requests to another provider or model. |
| Run on your own hardware | A local runtime or independent weights | You need Jev's calibration or accuracy: local runtimes wrap other models and were not compared on your task. |
| Open weights you can fine-tune or audit | Independent weights (Kev, CLM, Laya, Jeff…) | You cannot check separate code, adapter and base-model licences, or you need more than the model's context (Laya about 512 tokens, CLM 2k by default). |
| A hosted decision API, but not Jev | Another provider's hosted decision API (OpenAI Decisions API, generally available since 22 Oct 2026; Microsoft-Decision-1; Cloudflare Clef family; SubQ Snap, public preview; Liquid AI d1-ultra; Mouse-1, preview; Eevo; Drex (Nace.AI)) | Your data must not go to that provider, or you need Jev's exact answers (a different model answers). |
| Free-text output, or you already have labelled data | An LLM with structured output, or a trained classifier (10 criteria compared) | You need per-option probabilities without training. |
All alternatives, same columns
| Project and source | Layer replaced | Licences (code / weights / base) | Runtime and API shape | Parity or quality evidence | Checked (2026) |
|---|---|---|---|---|---|
| Kev · Kev-4B card | Independent weights | Code and weights Apache-2.0 (repo and cards); bases Qwen3.5-4B-Base and Qwen3.8-27B are Apache-2.0 (Hugging Face metadata); Kev-9B base not checked | Local or server; System One-compatible API; English only. Same rows as Jev and Laya | Author-run benchmarks; one operator-run comparison (Opper, below; graded B11) | Checked 2 Nov 2026: latest release still kev-1.0, not yet read. 31 Oct 2026: new release kev-1.0 (1 Oct 2026), not yet read; the row describes Kev as checked on 1 Oct 2026, 03:58 UTC, before that release (R25-S262). Earlier: 1 Oct: repository pushed 1 Oct, 03:58 UTC; Kev-9B and Kev-27B cards changed 30 Sep; no new GitHub release since 20 Sep. A third-party CPU ONNX server (kev-onnx) appeared 27 Sep, not audited |
| CLM | Independent weights | Code and CLM-8B weights Apache-2.0 (repo); Qwen3-8B base | Local playground; POST /v1/systemone with Choice, Noul, Score. Default 2k context, 8k tested | Author-run parity and speed claims | 25 Sep |
| Laya card · implementation | Independent weights | Code and weights Apache-2.0. The card metadata names no base model; the card's table names ModernBERT-large, and the multilingual checkpoint is mmBERT-base (MIT on Hugging Face) | Local decision head, router and server with runtime-defined options. Variants differ: the card model is a ModernBERT-large head (about 512 tokens); the README's multilingual and typed-decision variants use mmBERT-base (1,024 default context). The multilingual card says it "ships uncalibrated". Same rows as Jev and Kev | Card reports its own accuracy, calibration and many-option limits | 7 Nov 2026: Laya v0.4.28 (7 Nov 2026) is the latest release; contents not read; the row describes v0.3.22 (R27-S35). Earlier: 2 Nov 2026: latest GitHub release v0.4.21; PyPI laya 0.4.19, 0.4.20 and 0.4.21 published since the 31 Oct check, the last on 1 Nov 2026, 21:16 UTC; new release, not yet read; the row describes v0.3.22 (R26-S33, R26-S118). Earlier: 31 Oct 2026: new release v0.4.18 (30 Oct 2026), not yet read; the row describes v0.3.22 (R25-S263). Earlier: 1 Oct: code release v0.3.22 (29 Sep, 17:48 UTC), no new weights; cards unchanged since 24 Sep |
| Jeff (Gestalt-Lab) | Independent weights | Code and adapter Apache-2.0 (README); Qwen3-4B base terms: check | LoRA adapter; typed Choice, Noul, Score read from token probabilities | Author-run fact-check comparison | 25 Sep |
| Open-Jev (Zefan-Cai) | Independent weights | Not recorded; Qwen bases downloaded separately | Decision head; repo says 2B, 9B and 27B adapters published; local API | Author-run, version-specific benchmark | 25 Sep |
| system-one (mpuig) | Independent weights | Repo MIT; artifact and base terms: check | Qwen3-0.6B adapter and MiniCPM5-2B-q8 artifacts; Apple MLX server | 24-case comparison; not parity evidence | 25 Sep |
| NanoJev | Independent weights | Code MIT; Qwen base terms: check | 0.6B game-control checkpoint; zero-token readout | Author-run game tasks | 25 Sep |
| vjev-vision | Independent weights | Not recorded; Qwen3.5-4B-derived | Local image-state service (image input; Jev was text-only when this row was checked on 25 Sep; since 14 Oct 2026 Jev takes images only as an experimental opt-in, not production-ready per TypeSafe; since at least 7 Nov 2026 (first observed 7 Nov, 15:29 UTC; the change happened after 2 Nov, 11:13 UTC) TypeSafe's models page lists the input of both Jev 1.14 and Jev 1.13 as "Text, images" and says Jev "also reads PNG and JPEG images anywhere in the request"; the images page still calls images and Box "not yet ... ready for production" and requires the TypeSafe-Danger-Zone-Images: true header (R27-S05, R27-S06)) | Card reports ECE 0.183, option-order sensitivity, 512px resizing | 25 Sep |
| JevAny | Independent weights | Repo code claims Apache-2.0; adapter files and licences unverified; Qwen3.8-27B base Apache-2.0 | Two rank-16 LoRA checkpoints claimed; other tracks are prototypes | None | 26 Sep |
Strands Decider 2B (strands-labs) (0ac22a9). Copies are published by Amazon's verified Hugging Face organisation (Documented); The New Stack reports it as an AWS project (Reported). Amazon copy · The New Stack, 1 Oct | Independent weights; also an interface and a local runtime substitute | Code Apache-2.0 (repo); weights apache-2.0 (card metadata and LICENSE.md, StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 at 216d1cc); base Qwen/Qwen3.5-2B-Base apache-2.0 (Hugging Face metadata). Gemma 4 variants: bases Apache-2.0 per the README only (metadata not read) | Local: strands-decider serve exposes localhost:8000/v1/systemone with state and questions; --vision lets a request carry base64 images on the Qwen model (README); RTX 3090, Apple-silicon Macs or CPU. Reference model now strands-decider-2B-qwen3.5-v1-2610, plus four Gemma 4 variants (E2B, E4B, 12B, 26B-A4B), all 9 Oct on Hugging Face; no tags or GitHub releases | Authors' own: v1 2B at 180/231 on JevBench (Brier 0.280) Reported. Not compared with Jev by this pool | 10 Oct, 05:12–05:13 UTC: HEAD 0ac22a9 (9 Oct, 23:33 UTC; was 75c9fd3); repository created 29 Sep |
Jiwo 0.8B (52d217f) · code (8716aba) | Independent weights; also a local runtime | Code MIT (GitHub licence detection); weights apache-2.0 (card metadata); base Qwen/Qwen3.5-0.8B apache-2.0 (Hugging Face metadata) | jiwo serve on 127.0.0.1:8765 with a Jev-shaped POST /v1/systemone, plus GET /v1/models and /health; CUDA, MPS or CPU (Python 3.12). Text input only (no image input mentioned). Choice (up to 255 options), score (2–10 levels) and noul; training context 8,192 tokens. The card says "jiwo is an independent project" and that TypeSafe does not endorse it | Author's own: Decision Index 0.3 public 29.37 (v2); "1st of 27 models under 1B"; median 14.5 ms on one RTX PRO 6000 Reported. Not compared with Jev by this pool | 10 Oct, 05:12–05:13 UTC |
TOD (Parsec) (07d34d5) · code (642834e) · hosted page | Independent weights; also a hosted API from another provider (counted once, under weights) | Weights (LoRA adapter plus a ModernBERT-base retriever) apache-2.0 (Hugging Face metadata). Code: no licence detected by GitHub; the README puts the adapter and retriever under the bundle LICENSE. Base google/gemma-4-12B-it: not recorded, because the sources conflict (Hugging Face README: Apache-2.0; GitHub README: "governed by the Gemma 4 terms" plus the Prohibited Use Policy); base card not read; base not included in the download | Local inference.py CLI. Hosted API at $0.30 per 1,000 decisions (vendor page; price Documented); hosted API shape not read; no /v1/systemone in the GitHub README. Text plus images. Context 48,000 tokens (page) or max_tokens 49,152 (card). Same GitHub organisation as the daseinlabs/open-jev runtime row | Author's own: public JevBench 0.844; raw ECE 0.008 Reported. The vendor page says "~400M parameters", but the card is a 12B Gemma LoRA | 10 Oct, 05:13 UTC |
| Cloudflare Clef weights on Hugging Face: Cloudflare/clef, Cloudflare/clef-flash (both created 30 Sep 2026), Cloudflare/clef-omni (created 9 Oct 2026) (Cloudflare) | Independent weights: a different trained model; answers differ from Jev | Code: no separate code licence found; the repos ship custom code under the repo LICENSE, Apache-2.0. Weights: Apache-2.0 (Hugging Face tag and LICENSE file). Bases: Qwen3.8-27B and Qwen3.5-9B Apache-2.0 (card and LICENSE file); Qwen3-Omni-30B-A3B-Instruct card "license: other / license_name: apache-2.0", no LICENSE file in the repo (13 Oct 2026, 05:18 UTC; R15-S219 to R15-S221). Weights licence: R14-S406 | Cloudflare's changelog calls the Clef family "open-weight decision models"; it says the Clef-flash weights on Hugging Face are unchanged by the hosted context cut Reported (R14-S405). Cloudflare describes Clef as a "27B multimodal decision model", Clef-flash as a "fast 9B multimodal decision model" and Clef-omni as a "30B-parameter mixture-of-experts (3B active) backbone" (R14-S400–S402). Local serving requirements: not read. The game demo ra2web calls a local clef-flash by default (R14-S106); see its row on the real-time control page | As the hosted row: B32 used the hosted models; the weights themselves have no graded comparison. Not compared with Jev by this pool | 12 Oct, 05:34 UTC; base licences 13 Oct, 05:18 UTC |
| Ollaya · HN, 609 points | Local runtime or research code | Code Apache-2.0; each model keeps its own licence (weights pulled from the authors' repositories, not re-hosted); not checked model by model | Local daemon serving third-party decision models (Laya, decider, Kev, Winnow, CLM, GLiClass, NLI and others) with ONNX Runtime and llama.cpp; author documents POST /v1/systemone, /v1/decisions, GET /v1/models. v0.7.4 and v0.7.5 released 28 Sep (v0.7.5 adds image input) | Author-run, e.g. winnow:e4b "0.722 … (Jev: 0.738)" on LocalLLaMA/typed-decisions, n not stated (graded D, B19). API compatibility not tested by the pool | 28 Sep |
| OpenDecision | Local runtime or research code | Apache-2.0 (licence); upstream model terms: check | Local ModernBERT-large-zeroshot runtime; System One-shaped endpoint; v0.1.1 developer preview | README calls scores uncalibrated | 25 Sep |
| LocalJev | Local runtime or research code | Not recorded | Local DiffusionGemma/oMLX bridge exposing /v1/systemone | Probabilities are generated JSON, not a logit readout | 25 Sep |
| reflex | Local runtime or research code | Not recorded | Qwen3.5 prompted or LoRA runtime | Project documents weaker transfer for its trained adapter | 25 Sep |
| Jeff / GliFormer (logan-markewich) | Local runtime or research code | Not recorded | 400M GLiFormer-backed API (not Gestalt-Lab Jeff) | Approximate probability and confidence | 25 Sep |
| daseinlabs/open-jev | Local runtime or research code | Code MIT; Gemma licence separate | Gemma 3 4B option log-probability scoring; local endpoint | Doom demo uses text-described frames, not native vision | 25 Sep |
| open-alternative-jev | Local runtime or research code | Not recorded | In-process scorer using a user-supplied LLM; no HTTP endpoint | Author benchmark; packed batching can reduce accuracy | 25 Sep |
| AnyJev | Local runtime or research code | Not recorded | Research code for raw or corrected logits and fitted heads; needs labels and base weights | Author calibration and latency; not parity | 25 Sep |
| GPT-AGI/OpenJev | Local runtime or research code | Not recorded | Local readout and REPL; quick start is a mock | Trained head, weights and endpoint not shipped yet | 25 Sep |
| GemmaJev | Local runtime or research code | Not recorded | Local Gemma 4 option readout; separate media path | README disclaims reproduction and warns of bugs | 25 Sep |
| sgoedecke/system-one | Local runtime or research code | Not recorded | Local stock-Qwen single-token choice readout; a demo, not a trained model | None | 25 Sep |
| asynq-io/system-one | API interface adapter | Not recorded | Provider-neutral Python contract; local ONNX path using convaiinnovations/laya | Not applicable (interface) | 25 Sep |
| Jevper (PyPI 0.7.11, 26 Sep) | API interface adapter | Apache-2.0; no weights | In-process Python: a Jev-shaped system_one() over any OpenAI-compatible model, using log-probabilities or the model's JSON self-report; can also call hosted Jev, Ollaya and CLM. States it is "not affiliated with, endorsed by, or supported by TypeSafe AI" | None. With a reasoning model it reads a self-report, not a real distribution | 28 Sep |
| iamaamir/system-one | API interface adapter | Not recorded | Provider-neutral TypeScript builders | Not applicable (interface) | 25 Sep |
| TypeSafe System One adapter | API interface adapter | Not recorded | Adapter backed by other LLM APIs; data goes to those providers | Not applicable (interface) | 22 Sep, not rechecked |
| OpenAI Decisions API (generally available since 22 Oct 2026) Compare side by side: Decisions API vs Jev vs Microsoft Decision 1 | Hosted API from another provider: the whole hosted service. A different provider, model and request shape (POST /v1/decisions, not /v1/systemone). Not a product built on Jev | Not applicable (hosted, proprietary) | OpenAI's Decisions API became generally available on 22 Oct 2026 (OpenAI API changelog; guide read 23 Oct 2026, 05:23 UTC). gpt-6-luna is still the only model and input still costs $0.10 per 1M tokens; requests may now carry up to 400 questions (200 during the beta) (R22-S32, R22-S47). Earlier: public beta, "we expect to GA in the coming weeks" (13 Oct 2026; R15-S213). Input: text, images and JSON: "The Decisions API evaluates text, images, and JSON" (guide, checked 31 Oct 2026, 05:11 UTC; R25-S206; checked 2 Nov 2026, unchanged; checked 7 Nov 2026, 15:36 UTC, unchanged, R27-S27) (Jev's models page said "Text only" until at least 14 Oct 2026, 16:46 UTC; since 14 Oct, Jev takes images only as an experimental opt-in, not production-ready per TypeSafe; 16 Oct 2026, 05:22 UTC; R17-S05, R17-S06; since at least 7 Nov 2026 (first observed 7 Nov, 15:29 UTC; the change happened after 2 Nov, 11:13 UTC) TypeSafe's models page lists the input of both Jev 1.14 and Jev 1.13 as "Text, images" and says Jev "also reads PNG and JPEG images anywhere in the request"; the images page still calls images and Box "not yet ... ready for production" and requires the TypeSafe-Danger-Zone-Images: true header (R27-S05, R27-S06)). A per-request image limit was not found in the guide, API reference or changelog on 30 or 31 Oct 2026 (R25-S206 to R25-S210); the 13 Oct 2026 reading showed up to 128 images per request. Change, 31 Oct 2026: the guide's input wording changed between 13 and 31 Oct 2026; earlier: a text string, or user messages with text and images; "at most 128 images are allowed across all messages in one request" (API reference, 13 Oct 2026; R15-S214). Answers: predicate, choice and score. Refusal: "The model declined to answer this question. Other questions in the same request can still receive answers." (API reference, 13 Oct 2026; R15-S214; unchanged at 31 Oct 2026, 05:11–05:12 UTC, R25-S208, R25-S209). Error codes, timeouts and retries: not found in the guide, the API reference or the error-codes guide (13 Oct 2026, 05:18 UTC; R15-S213, R15-S214, R15-S216). Price as published: $0.10 per 1M input tokens; output not charged; since GA (22 Oct 2026) cache reads are discounted ($0.01 per 1M cached input tokens, Standard tier, short context) and cache writes are charged ($0.125 per 1M for 30 minutes, $0.25 per 1M for 12 hours) (OpenAI guide and pricing page, 24 Oct 2026, 05:15 UTC; R23-S32, R23-S45, R23-S46; unchanged at 30 Oct 2026, 05:17 UTC, R24-S200, R24-S201); regional and long-context multipliers apply. In the public beta (13 Oct) there were no cache charges. ZDR and HIPAA eligible, per the guide. Regions: the guide says "Data residency is available in all supported API regions. Regional processing is supported in the United States and Europe (EEA + Switzerland)." (checked 31 Oct 2026, 05:11 UTC; R25-S206) Change, 31 Oct 2026: changed between 13 and 31 Oct 2026; earlier: "Data residency and regional processing are supported in the United States and Europe (EEA + Switzerland)" (R14-S219). OpenAI's data-controls regional table lists /v1/decisions for the United States, Europe, Australia, Canada, Japan, India and Singapore (13 Oct 2026; R15-S217; not re-read). Both shown; not merged. Routes: OpenAI direct; OpenRouter openai/gpt-6-luna-decisions (13 Oct 2026; R15-S49); Vercel AI Gateway openai/gpt-6-luna-decisions (13 Oct 2026; R15-S212); both unchanged at 31 Oct 2026, 05:14–05:19 UTC (R25-S42, R25-S43); checked 2 Nov 2026, unchanged. Rechecked 13 Oct 2026: status, model, price and data controls unchanged (05:18 UTC; R15-S213, R15-S217). Earlier check: 12 Oct 2026, 05:13 UTC (R14-S219). OpenAI's data-controls table gives /v1/decisions: data used for training "No"; abuse monitoring retention "30 days"; ZDR eligible "Yes, see below for limitations"; Private Retention "Pending confirmation" (undated page, read 12 Oct 2026, 05:13 UTC, R14-S218; unchanged 13 Oct 2026, 05:18 UTC, R15-S217). Change, 30 Oct 2026: OpenAI now lists the Decisions API as supporting Zero Data Retention, Private Retention with PSP and Safety Retention for eligible customers ("Private Retention with PSP and Safety Retention eligible: Yes, see below for limitations"); abuse-monitoring logs are kept for up to 30 days by default; prompt caching may keep encrypted key/value tensors until a 24-hour expiry (OpenAI "Your data" page, read 30 Oct 2026, 05:17 UTC; R24-S203). The entry changed between 13 Oct 2026, 05:18 UTC and 30 Oct 2026, 05:17 UTC (the page gives no date for it); the "Pending confirmation" value above is kept as dated. The /v1/decisions row (training "No"; abuse monitoring "30 days"; ZDR and Private Retention with PSP and Safety Retention "Yes, see below for limitations") is unchanged at 31 Oct 2026, 05:17 UTC (R25-S242). OpenAI's pricing page has no Decisions API line; the $0.10 figure is published only in the guide (12 Oct 2026; R14-S217). On 24 Oct 2026 (05:15 UTC) the pricing page lists gpt-6-luna with its input, cached-input and cache-write prices, which the guide says Decisions caching matches (R23-S45, R23-S32). Your data goes to OpenAI (Jev's route data terms do not apply) | Six graded comparisons with Jev (24 Oct 2026; five until then); results as their operators report them Reported: RPG NPC move choices: the Decisions API came out ahead (82 of 100 matches; per-move accuracy 59.4% vs 51.0%) (B38, A, komzweb; flags: game task, operator-designed task, one run, full call logs not committed). Portuguese routing, injection detection and judging: Jev ahead on exact judge score (80.7% vs 72.0%); the Decisions API ahead on judge score within ±1 point (100.0% vs 98.7%) and on injection precision (100.0% vs 97.3%); the operator reports no significant difference on routing accuracy (88.0% vs 86.7%) or injection F1 (96.6% vs 95.8%) (B39, A, mateusfonsek; flags: single language, LLM-generated dataset, one run per case, Jev called through the jev-latest alias, p-value source not located). Synthetic bank-withdrawal routing: On synthetic bank-withdrawal routing (B41, grade A), Jev (89.2% macro-F1) and Microsoft-Decision-1 (87.6%) were not separated and both were ahead of OpenAI's Decisions API (71.4%); Decision-1 kept the same route across all 6 option orders more often (97.2–97.6% vs Jev 91.6–91.8%, Decisions API 64.2%). (B41, A, kartikanand73/decision-api-benchmark at 1ffa4d5, 500 synthetic cases × 6 option orders, two runs on 10 Oct 2026; flags: synthetic single task; thresholds set with Jev-shaped probabilities, the operator's own note; all three models called via OpenRouter; code version not pinned; checked 12 Oct 2026, 05:12 UTC; R14-S201–S206). Five-option logic questions: On 1,000 five-option logic questions (B42, grade B), Jev answered 817 correctly and OpenAI's Decisions API 612; Jev's answers come from a 22 Sep run, and the answer key is not published. (B42, B, GitHub account etsabary, etsabary/luna-decisions-benchmark at 7a5ea13, 1,000 five-option deterministic logic questions (25 families); Decisions API called as openai/gpt-6-luna-decisions via OpenRouter; flags: via OpenRouter's alpha Decisions endpoint, not OpenAI direct; Jev not re-run (22 Sep vs 10 Oct; mixed Jev providers, 256 answers without a verified snapshot); answer key withheld (only its sha256 committed); 17 persistent refusals (HTTP 502), cause unknown; the operator built both benchmarks; young repository (2 commits; HEAD changed after the main publish commit); latency and cost not comparable across dates and routes; re-pin checked 13 Oct 2026, 05:24 UTC; R14-S300 to R14-S305, R15-S53). Tetris round robin: In a 48-game Tetris round robin (metric: games won; 8 games per pair), Microsoft-Decision-1 won 15, Jev 12, OpenAI's Decisions API 11 and the open Laya model 8, with 2 draws; Jev beat the Decisions API 6–2 head to head (B44, A, dante01yoon/decision-model-arena at 285c803; flags: small-n (8 games per pair); single task (games); Decision-1 through OpenRouter's alpha endpoint, not Foundry; Jev called as the jev-latest alias; one date and region (called from the US); files opened 14 Oct 2026, 16:55 UTC; R16-S67). Five public classification tests: B45 (gbolask24, grade A; Reported): on five public-data tests run 17 Oct 2026, accuracy showed no detectable difference on three and Jev 1.14 ahead on two (CLINC150, Decision Bench bench-v4); Brier: Jev better on all five. B45 ran on 17 Oct 2026, while OpenAI's Decisions API was in public beta (generally available from 22 Oct 2026) (B45, A, gbolask24/decision-bench at e5b7202, an individual, no commercial stake found; jev-1.14.0 vs gpt-6-luna, 7,070 items, one test with product photos (T3); flags: public-data; small-n (Opus arm); note: per-call logs referenced in the README are not in the repository; files opened 24 Oct 2026, 05:15–05:21 UTC; R23-S200 to R23-S211). Results differ by task; the pool states no overall parity. One more comparison, classifier-bakeoff, is a lead, not graded (report opened, metric code not opened). All leads | Evidence: B45 24 Oct, 05:15–05:21 UTC (graded at pinned commit); B44 14 Oct, 16:55 UTC (graded at pinned commit); B42 13 Oct, 05:24 UTC (re-pin; files opened 12 Oct); B41 12 Oct, 05:12 UTC; B38 and B39 11 Oct, 05:23 UTC (graded at pinned commits). Service details and data controls: 13 Oct, 05:18 UTC (status, model, price and data controls unchanged; refusal, image limit, regions and routes added). Status, model, price and question limit: 23 Oct, 05:23–05:25 UTC (R22-S32, R22-S47). Caching prices: 24 Oct, 05:15 UTC (R23-S32, R23-S45, R23-S46). Status, model, price and question limit rechecked, unchanged; data controls rechecked, retention wording changed: 30 Oct, 05:17 UTC (R24-S200 to R24-S203). Input and region wording changed, image limit not found, refusal, data controls and routes unchanged: 31 Oct, 05:11–05:19 UTC (R25-S206 to R25-S210, R25-S242, R25-S42, R25-S43) |
| Microsoft-Decision-1 (Microsoft; published 9 Oct 2026) · Foundry catalog · Microsoft Learn how-to · OpenRouter · HN item 50024913 (197 points; R13-S33) Microsoft's catalog: "built on Alibaba's open-weight Qwen3.5-9B and post-trained by Microsoft" (R14-S209) Compare side by side: Decisions API vs Jev vs Microsoft Decision 1 | Hosted API from another provider: replaces Jev's hosted API layer. A different provider and model. Your data goes to Microsoft under Microsoft's terms; through OpenRouter, also under OpenRouter's terms (not read). It is an alternative, not a product built on Jev | Not applicable (hosted service). Catalog licence text: "Use of this Azure Direct Model is subject to supplemental Terms of Use for Microsoft Azure" (supplemental terms; model page, 13 Oct 2026, 05:20 UTC; R15-S44; model page, checked 2 Nov 2026, 11:00 UTC, unchanged; R26-S41). Microsoft's Product Terms entry "Foundry Models sold by Azure" applies to models sold by Azure; it does not name Decision-1 (Effective 10/1/2026 view, 13 Oct 2026; R15-S200). The model-specific terms page has no Decision-1 or Qwen entry (R15-S202) | Documented by Microsoft (Learn, Last updated on 2026-10-09; read 12 Oct 2026, 05:13 UTC; R14-S208; Learn, checked 7 Nov 2026, 15:36 UTC, same date, R27-S34). Request: POST <your-foundry-resource-endpoint>/providers/microsoft/v1/systemone, with Microsoft Entra ID (Bearer) or an api-key header; body model (your deployment name), state (text or JSON) and named questions of type noul, choice or score, with instructions and criteria. noul returns "A probability from 0 through 1"; choice "One selected option and the probability for each option"; score "A value on an ordered scale and the probability for each level". No separate confidence field is documented on that page. Request path, auth and the three answer types unchanged at 31 Oct 2026, 05:11–05:20 UTC (R25-S205, R25-S245); request path and auth checked 2 Nov 2026, unchanged. Model ID Microsoft-Decision-1, version 1 in Foundry; microsoft/microsoft-decision-1 on OpenRouter (served microsoft/microsoft-decision-1-20261009; R14-S212). Input: text only: "The model is text-only and does not accept or produce image, audio, or video content." (catalog, 12 Oct 2026; R14-S209; Foundry catalog, checked 2 Nov 2026, 11:00 UTC, unchanged; R26-S41). Limits: 32,768-token context (Foundry catalog and OpenRouter, checked 2 Nov 2026, 10:54–11:00 UTC; unchanged; R26-S41, R26-S35; OpenRouter's Azure endpoint 32,768 unchanged at 7 Nov 2026, 15:34 UTC, R27-S28). Earlier: OpenRouter's Azure endpoint unchanged at 31 Oct 2026, 05:20 UTC (R25-S246). Microsoft's quotas page: "The following table lists tokens per minute (TPM) and requests per minute (RPM) limits for Microsoft-Decision-1 by subscription tier. Your deployment's rate limits depend on its configured capacity and allocated quota." Tier 1 2.16M TPM / 2,160 RPM; Tier 2 4.50M / 4,500; Tier 3 6.00M / 6,000; Tier 4 9.00M / 9,000; Tier 5 10.20M / 10,200; Tier 6 15.00M / 15,000; the table no longer splits Global Standard and Data Zone Standard (Microsoft quotas page, checked 31 Oct 2026, 05:20–05:22 UTC; R25-S244, R25-S46). Change, 31 Oct 2026: changed between 13 Oct and 31 Oct 2026; Microsoft's page shows no new update date. Earlier: requests per minute by subscription tier, the same for Global Standard and Data Zone Standard: Tier 1 60, Tier 2 150, Tier 3 225, Tier 4 360, Tier 5 500, Tier 6 900; no tokens-per-minute figure (Foundry quotas page, Last updated 2026-10-06; 13 Oct 2026, 05:18 UTC; R15-S206). Learn documents "429 Too Many Requests … Retry with exponential backoff, or request more quota" (unchanged at 31 Oct 2026, 05:20 UTC; R25-S245). Deployment: "DataZoneStandard in selected regions; GlobalStandard" (R14-S208). Foundry lifecycle: listed as Preview in Foundry at 13 Oct 2026, 05:20 UTC (listed Generally available on 12 Oct; cause unknown; R15-S44, R14-S44); still Preview, version 1, at 16 Oct 2026, 05:21 UTC (R17-S33) and at 2 Nov 2026, 11:00 UTC (R26-S41). OpenRouter lists one provider, Azure (R14-S212); Vercel AI Gateway lists microsoft/microsoft-decision-1 (input 0.000000042, output 0; 13 Oct 2026; R15-S212). OpenRouter (one provider, Azure) and the Vercel listing unchanged at 31 Oct 2026, 05:14–05:19 UTC (R25-S42, R25-S43) and at 2 Nov 2026, 10:54 UTC (R26-S35, R26-S27). Price as published: "Input tokens cost $0.042 USD per million tokens. Output tokens are free." (Microsoft's post, rechecked 12 Oct 2026, 05:13 UTC, unchanged; R14-S207; unchanged at 30 Oct 2026, 05:19 UTC, R24-S208; Microsoft's post, checked 7 Nov 2026, 15:36 UTC, unchanged, R27-S34; OpenRouter input 0.000000042, one provider, Azure, 30 Oct 2026, 05:19 UTC, R24-S209); OpenRouter: "$0.042/M input tokens and $0.00/M output tokens" (R14-S213). No Microsoft price page was read; the catalog's "View pricing" needs sign-in. Data terms (partial): Microsoft: prompts and completions "are not used to train, retrain, or improve the base models"; Microsoft stores and processes data "to monitor for uses that violate the applicable product terms" ("Data, privacy, and security for Foundry Models sold by Azure", Last updated on 2026-05-19; read 12 Oct 2026; R14-S215). No retention period stated on the pages read (12 Oct 2026, 05:13 UTC); the page does not name Decision-1, which Microsoft lists as a model sold by Azure (R14-S210). Microsoft's Product Terms ("Foundry Models sold by Azure", Effective 10/1/2026 view): "Microsoft will temporarily store Input and Output Content, to monitor for and prevent abusive or harmful uses or outputs of the service." No period in days on the Product Terms, the model-specific terms, the data page or the abuse-monitoring page (13 Oct 2026, 05:18–05:20 UTC; R15-S200 to R15-S204). Unchanged at 31 Oct 2026, 05:11 UTC: the no-training sentence, the same "temporarily store" sentence on the Product Terms page as served (latest effective-date view listed: 10/28/2026), and still no period in days and no mention of Decision-1 on those pages or the Decision-1 how-to (R25-S200 to R25-S205). The no-training and "temporarily store" sentences checked 2 Nov 2026, unchanged. Vendor claims Reported: Microsoft says Decision-1 "changes its decision on 1.3% of perturbations on average with zero flips when option descriptions are paraphrased or when options are reversed or shuffled", and that its "P50 latency is ~35x faster than GPT-6 Sol" (R14-S207). New since 11 Oct: the post carries "Editor's note: This post was updated from the original to add benchmarks for Jev on accuracy and calibration." The update date is not shown. Microsoft's charts in the post (decoded 13 Oct 2026, 05:18 UTC; R15-S209; Microsoft's own measurement, not graded by the pool): accuracy, "mean score across 36 benchmarks (147,137 questions)": Microsoft-Decision-1 83.5, Jev 1.13.0 82.3, GPT-6 Luna Decisions 79.4; calibration ("100 means a model's confidence exactly matches how often it is right"): Jev 1.13.0 93.7, Microsoft-Decision-1 92.2, GPT-6 Luna Decisions 89.9; median latency (JevBench v1.6.1 adjusted p50, "checked Oct 7, 2026"; Decision-1 "measured through Foundry in the same region"): Decision-1 85 ms, Jev 1.13.0 240 ms, GPT-6 Luna Decisions 300 ms | Two graded comparisons with Jev; results as their operators report them Reported: Synthetic bank-withdrawal routing: On synthetic bank-withdrawal routing (B41, grade A), Jev (89.2% macro-F1) and Microsoft-Decision-1 (87.6%) were not separated and both were ahead of OpenAI's Decisions API (71.4%); Decision-1 kept the same route across all 6 option orders more often (97.2–97.6% vs Jev 91.6–91.8%, Decisions API 64.2%). (B41, A, kartikanand73/decision-api-benchmark at 1ffa4d5, 500 synthetic cases × 6 option orders, two runs on 10 Oct 2026; flags: synthetic single task; thresholds set with Jev-shaped probabilities, the operator's own note; all three models called via OpenRouter; code version not pinned; checked 12 Oct 2026, 05:12 UTC; R14-S201–S206). Tetris round robin: In a 48-game Tetris round robin (metric: games won; 8 games per pair), Microsoft-Decision-1 won 15, Jev 12, OpenAI's Decisions API 11 and the open Laya model 8, with 2 draws; Jev beat the Decisions API 6–2 head to head (B44, A, dante01yoon/decision-model-arena at 285c803; flags: small-n (8 games per pair); single task (games); Decision-1 through OpenRouter's alpha endpoint, not Foundry; Jev called as the jev-latest alias; one date and region (called from the US); files opened 14 Oct 2026, 16:55 UTC; R16-S67). Two tasks; the pool states no overall parity. Not compared with Jev by this pool | 7 Nov, 15:34–15:36 UTC (Microsoft's post price, Learn date and OpenRouter endpoint only; unchanged; R27-S34, R27-S28). 2 Nov, 10:54–11:00 UTC (Foundry catalog text-only input, 32,768-token context, licence text and Preview lifecycle re-read; request, data terms, OpenRouter context and routes rechecked; all unchanged; R26-S41, R26-S35, R26-S27). 31 Oct, 05:11–05:22 UTC (quotas page: rate limits changed; request, answer types, 429 line, data terms, Product Terms, OpenRouter context and routes rechecked and unchanged; catalog text, licence and lifecycle not re-read; R25-S200 to R25-S205, R25-S244 to R25-S246, R25-S46, R25-S42, R25-S43). 30 Oct, 05:19 UTC (price, rechecked and unchanged; R24-S208, R24-S209). 16 Oct, 05:21 UTC (Foundry lifecycle). 14 Oct, 16:55 UTC (B44 files). 13 Oct, 05:18–05:20 UTC (quotas page, Product Terms, post charts and Vercel 05:18–05:19; catalog lifecycle and licence 05:20). 12 Oct, 05:12–05:15 UTC (B41 05:12; Microsoft post, Learn, catalog and OpenRouter 05:13–05:14; data page 05:13). HN item: 11 Oct, 05:21 UTC |
Cloudflare Clef family on Workers AI: Clef, Clef-flash, Clef-omni (Cloudflare; model IDs @cf/cloudflare/clef, @cf/cloudflare/clef-flash, @cf/cloudflare/clef-omni; Clef-omni added 9 Oct 2026)Compare side by side: Decisions API vs Jev vs Microsoft Decision 1 | Hosted API from another provider: replaces Jev's hosted API layer. A different provider and model; your data goes to Cloudflare under Cloudflare's terms on Workers AI; on OpenRouter's non-Cloudflare hosts (Clef: DeepInfra, PrimeIntellect; Clef-flash: Makora, DeepInfra, PrimeIntellect; checked 31 Oct 2026) it goes to that host. It is an alternative, not a product built on Jev. (This is a different listing from Jev's own Cloudflare route typesafe/jev, which serves TypeSafe's model; see versions and aliases.) | Not applicable to the hosted service; the weights are listed separately (weights row) | Documented by Cloudflare (R14-S400–S402). Request: the same state-plus-typed-questions shape as Jev (noul, choice, score), through env.AI.run or Cloudflare's REST /ai/run/@cf/cloudflare/<model> with a Bearer token, with Cloudflare extensions for images (all three) and audio and video (Clef-omni). The endpoint, auth and extension fields differ from Jev's, so a Jev client does not work unchanged. Context: Clef 65,536 tokens; Clef-flash 24,576 (Cloudflare's changelog: "now 24K tokens, down from 64K"; R14-S405); Clef-omni 64,000. Price on Cloudflare's Workers AI pricing page (12 Oct 2026, 05:33 UTC; R14-S403): Clef "$0.240 per M input tokens", Clef-flash "$0.038 per M input tokens", Clef-omni "$0.150 per M input tokens"; "Clef models do not charge for output tokens". Prices and contexts unchanged at 30 Oct 2026, 05:17 UTC; the pricing page now shows "Last updated Oct 20, 2026" (R24-S216 to R24-S219). Prices and contexts unchanged on the three model pages at 31 Oct 2026, 05:27 UTC (R25-S48; pricing page not re-read). Data: training: "No training: Cloudflare does not use Customer Content processed by Workers AI to train AI models." Retention: "Zero data retention: Workers AI processes Customer Content only for as long as needed to provide the service. For standard inference requests that do not invoke a stateful feature, Cloudflare does not retain inputs or outputs after they are no longer needed to process the request and deliver the response, subject to the limited exceptions below." Exceptions include: "Customer Content may be stored when you: use a Cloudflare storage service, such as R2, KV, Durable Objects, or Vectorize, with Workers AI; enable logging or caching in AI Gateway or write Customer Content to application logs". No period in days stated (Data usage, Last updated Oct 16, 2026; checked 31 Oct 2026, 05:14 UTC; R25-S239, R25-S45). Cloudflare's data-usage page (updated 16 Oct 2026) says "Zero data retention: Workers AI processes Customer Content only for as long as needed to provide the service", with exceptions (storage services, AI Gateway logging or caching); no period in days stated. This covers Workers AI; OpenRouter's non-Cloudflare hosts are separate. Change, 31 Oct 2026: the page is dated 16 Oct 2026 and was first seen in this form on 31 Oct 2026. Earlier: Cloudflare "does not use your Customer Content to (1) train any AI models made available on Workers AI or (2) improve any Cloudflare or third-party services, and would not do so unless we received your explicit consent." (Data usage, Last updated Apr 21, 2026; read 12 Oct 2026, 05:33 UTC; R14-S404); no retention period stated on that page. The 12 Oct version also said "Cloudflare neither creates nor trains the AI models made available on Workers AI", which sat oddly with Cloudflare publishing Clef as its own model; whether the 16 Oct version still says this was not recorded. Errors: Workers AI's platform error table, not Clef-specific: "Timeout | 3007 | 408"; "Out of capacity | 3040 | 429 | Capacity temporarily exceeded, please try again." (errors page, Last updated Sep 17, 2026; 13 Oct 2026, 05:18 UTC; R15-S223; quoted rows unchanged at 31 Oct 2026, 05:14 UTC, R25-S241). A Clef-specific rate limit was not found on the Workers AI limits page at 13 Oct 2026 (R15-S222) or at 31 Oct 2026, 05:14 UTC (R25-S240). Other routes: also on OpenRouter, including non-Cloudflare hosts: cloudflare/clef via DeepInfra (65,537 context), PrimeIntellect (65,536) and Cloudflare (65,536); cloudflare/clef-flash via Makora (16,384), DeepInfra (24,576), PrimeIntellect (24,576) and Cloudflare (24,576); cloudflare/clef-omni via Cloudflare (OpenRouter endpoints, checked 31 Oct 2026, 05:19 UTC; R25-S42). On OpenRouter, Clef-flash is now also served by Makora and DeepInfra; on any non-Cloudflare host, your data goes to that host, not to Cloudflare (OpenRouter endpoints, checked 31 Oct 2026, 05:19 UTC). The pool read the host lists only, not the hosts' terms. Change, 31 Oct 2026: hosts and contexts changed between 13 Oct and 31 Oct 2026. Earlier: cloudflare/clef via PrimeIntellect (16,384 context), DeepInfra and Cloudflare; cloudflare/clef-flash via PrimeIntellect and Cloudflare; cloudflare/clef-omni via Cloudflare (13 Oct 2026, 05:21 UTC; R15-S226). Not listed in Vercel AI Gateway's model list at 13 Oct 2026, 05:18 UTC (R15-S212) or at 31 Oct 2026, 05:14 UTC (R25-S43) | B32 (A, system-one-security, an individual GitHub repository; security experiments on a synthetic citation-checker knowledge base; flags: small-n, no-baseline, sealed, single-task; graded 7 Oct 2026, 05:20 UTC) uses Clef and Clef-flash as comparators; its 2,048-token truncation finding concerns Clef, not Jev. Cloudflare's own benchmark table and its statement that Clef-flash "now costs less than Jev" are Reported (Cloudflare's changelog, 9 Oct 2026; R14-S405). Not compared with Jev by this pool; the pool made no per-decision cost comparison | 31 Oct, 05:14–05:27 UTC (data-usage page changed; OpenRouter hosts changed; model-page prices and contexts, errors page and limits page rechecked and unchanged; R25-S239, R25-S45, R25-S42, R25-S43, R25-S48, R25-S240, R25-S241). 30 Oct, 05:17 UTC (prices and contexts, rechecked and unchanged; R24-S216 to R24-S219). 13 Oct, 05:18–05:21 UTC (errors page, limits page, OpenRouter endpoints). 12 Oct, 05:33–05:34 UTC |
| SubQ Snap (Subquadratic) (public preview) Compare the hosted decision APIs: Decisions API vs Jev vs Microsoft Decision 1 (SubQ Snap is not a column there) | Hosted API from another provider: the whole hosted service. A different provider and model (subq-snap-preview at 14 Oct 2026; SubQ's docs list subq-snap-1.1 as the default Decision model at 30 Oct 2026, R24-S212, R24-S213) with the same request shape as Jev's (POST /v1/systemone: state plus named noul, choice or score questions). Not a product built on Jev | Not applicable (hosted, proprietary) | Documented by Subquadratic. Status and request: public preview: "public preview is open to every account" (docs, no date shown, read 14 Oct 2026, 16:47 UTC; R16-S217; docs llms.txt, checked 7 Nov 2026, 15:48 UTC, unchanged, R27-S47). The launch post adds "Once capacity is reached, new users may be added to a waitlist" (12 Oct 2026; R16-S224). Request: POST https://api.subq.ai/v1/systemone with "Authorization: Bearer <API key>"; body model (required), state, 1–16 named questions, optional images (up to 4); context "384K-token (393,216)" (decisions API docs, no date shown; R16-S218). Not tested: whether Jev's SDKs work against it by changing the base URL. Price as published: SubQ Snap 1.1 (subq-snap-1.1), public preview: SubQ's models page lists "Pricing: Input ($0.04 / 1M tokens), Output (Free)" (page modified 28 Oct 2026; read 30 Oct 2026, 05:17 UTC; R24-S212). The 12 Oct launch post still says "$0.05 per million input tokens" (read 30 Oct 2026, 05:17 UTC; R24-S214); both are shown, not reconciled. Both unchanged at 31 Oct 2026, 05:27 UTC (docs: subq-snap-1.1, "Input ($0.04 / 1M tokens), Output (Free)"; launch post "$0.05 per million input tokens"; R25-S47); checked 2 Nov 2026, unchanged. Docs price "Pricing: Input ($0.04 / 1M tokens), Output (Free)" checked 7 Nov 2026, 15:48 UTC, unchanged (llms-full.txt; R27-S47); the launch post price was not compared on 7 Nov. Change, between 14 Oct and 28 Oct 2026: model version and docs price. Earlier: "Pricing: Input ($0.05 / 1M tokens), Output (Free)" (docs models page, no date shown, read 14 Oct 2026; R16-S220). Same figure in the launch post of 12 Oct 2026 (R16-S224). There is no separate price page: subq.ai/pricing serves the home page (14 Oct 2026, 16:47 UTC; R16-S225). Requests need prepaid credits: "Make sure your account has credits before sending requests." (R16-S221). Your data: training: "We do not train on Customer Content. We do not use Customer Content to train or improve our foundational models." (Terms, Effective Date October 12, 2026; R16-S222). Retention: two texts that disagree, both shown and not merged. Terms: "We retain Input and Output, including related safety classification results, for no more than thirty (30) days after processing, after which they are deleted, unless longer retention is required by law…" (Effective Date October 12, 2026; R16-S222). Data-retention docs: "Retain indefinitely (default) — kept until you delete it." (no date shown, read 14 Oct 2026, 16:47 UTC; R16-S219). Zero data retention: "available on request and subject to our approval" (Terms; R16-S222). Your data goes to Subquadratic (Jev's route data terms do not apply). Routes: Subquadratic direct only. Not listed on OpenRouter (462 models) or Vercel AI Gateway (464 models) at 14 Oct 2026, 16:47 UTC (R16-S226, R16-S227). Errors and limits: documented error codes 400, 401, 402, 403, 422, 429 and 5xx, with "Retry 429 and 5xx with backoff." (R16-S218). Numeric rate limits: not found in the docs (all pages in llms-full.txt, the decisions API, quickstart, authentication, models and data-retention pages) or the Terms at 14 Oct 2026, 16:48 UTC | Two graded comparisons with Jev (24 Oct 2026; one until then); results as the operators report them Reported: SWE-bench Verified file localization (200 issues, 10 candidate files): SubQ Snap 78.5% and Jev 79.0% when picking one file; asked per file, Snap 85.5% vs Jev 81.8%, a difference the operator reports as not significant (95% CI −0.5 to +8.3 points) (B43, B, kprzybys/snap-vs-jev-swe at b754c44; flags: COI (prize challenge run by the competing vendor); single runs; preview model with a vendor-confirmed server change on 10 Oct; public data; post-hoc question wording on one dataset; young repository (1 commit); R16-S66). COI: the operator takes part in Subquadratic's Snap creator challenge, which has prizes. Single runs; preview model. Synthetic event-log QA: B46 (jcarlosrodicio/subq-vs-jev, grade A; Reported, 13 Oct 2026): synthetic event-log QA; accuracy up to 30K tokens Snap 93.9%, Jev 94.7%; at 360K Snap 58.8%, Jev 77.1% with the log split into 29K chunks; ECE up to 30K Snap 0.028, Jev 0.038 (B46, A, jcarlosrodicio/subq-vs-jev at c037905, an independent individual; jev-1.13.0 (called as jev-latest) vs subq-snap-preview, run 13 Oct 2026 in SubQ's public preview; the 29K chunking at 360K is the operator's; up to 30K the intervals overlap; flag: single-task; files opened 24 Oct 2026, 05:16 UTC; R23-S217 to R23-S223). Vendor claims against Jev (Subquadratic's own index) are listed separately and not graded (side-by-side page). Not compared with Jev by this pool | 7 Nov 2026, 15:36–15:48 UTC (docs llms.txt and llms-full.txt, launch post: status and price unchanged, "Input ($0.04 / 1M tokens), Output (Free)"; retention wording not determined, its earlier date kept; R27-S33, R27-S47). 2 Nov 2026 (models page and launch post: model version and prices checked, unchanged). 31 Oct, 05:27 UTC (models page and launch post: model version and prices rechecked and unchanged; R25-S47). 30 Oct, 05:17 UTC (models page, llms.txt and launch post: model version and price; R24-S212 to R24-S214). 14 Oct, 16:47–16:48 UTC (docs, Terms, launch post, home page, OpenRouter and Vercel model lists). B43 files: 14 Oct, 16:52 UTC. B46 files: 24 Oct, 05:16 UTC |
| Liquid AI d1-ultra (Liquid AI; launched 15 Oct 2026) · Liquid docs · pricing page Liquid's docs: "a larger open-weight model with Liquid-specific customizations" (base not named; R19-S101) | Hosted API from another provider: the whole hosted service. A different provider and model (d1-ultra) with the same request shape as Jev's (/systemone path: state plus named noul, choice or score questions) on its own endpoint. Not a product built on Jev | Not applicable (hosted, proprietary). d1-ultra weights not published; base model not named (docs, read 19 Oct 2026, 05:22 UTC; R19-S101). Not published; base model not named (docs, read 7 Nov 2026, 15:36 UTC; R27-S30) | Documented by Liquid AI. Status and request: "available today on the Liquid API" (launch post, 15 Oct 2026; R19-S100); no preview label found on the post, docs or pricing page (19 Oct 2026, 05:22–05:23 UTC; same at 31 Oct 2026, 05:11 UTC, R25-S211, R25-S213). Request: POST https://api.liquid.ai/decisions/v1/systemone with a Bearer key; body model, state, questions, optional images (up to 8, Base64 only, body under 4.5 MB; R19-S102; image count and body limit unchanged at 31 Oct 2026, R25-S212). Liquid's docs call it with TypeSafe's SDK by changing the base URL (text only: "the SDK does not support images yet"; R19-S102); not tested. Context 32,768 tokens (Vercel AI Gateway listing; R19-S28); no context figure found on Liquid's own pages. Errors and limits: error codes and rate limits not found on Liquid's decision-model docs (19 Oct 2026, 05:23 UTC; R19-S101 to R19-S103, R19-S111). Price as published: "d1-ultra Input $0.042 / MTok Cached input $0.042 / MTok Output $0.00 / MTok" (pricing page, no date shown, read 19 Oct 2026, 05:22 UTC; R19-S104; unchanged at 31 Oct 2026, 05:11 UTC, R25-S211; checked 7 Nov 2026, 15:36 UTC, unchanged, R27-S30). TypeSafe's published Jev price is $0.042 per million input tokens, output free (cost per decision). Images are billed as input at 1.5 tokens per 32×32-pixel patch (R19-S102), and "Each question is billed for its text and all supplied images" (docs, checked 31 Oct 2026, 05:11 UTC; R25-S212). Change, 31 Oct 2026: reworded between 19 Oct and 31 Oct 2026 (the page shows no date); meaning not interpreted. Earlier: "Each question is billed as its own prompt, including its text and all images." (R19-S102). A new organization gets "$2 in free credit" (R19-S104; unchanged at 31 Oct 2026, R25-S211). Your data: training: the Terms grant Liquid AI a licence "to use Your Content to develop and improve Liquid AI's models, products, and services, including through training, fine-tuning, evaluation, and testing" (Terms, Last updated SEP 30, 2026; R19-S105). Retention: "as reasonably necessary to provide, secure, and improve the Services; comply with law; resolve disputes; and enforce our agreements" (no period given; R19-S105). Opt-out, zero data retention and a DPA: not found on the Terms, Privacy Policy, pricing or docs pages (19 Oct 2026, 05:22–05:23 UTC). The pricing page's "we never use customer data to train models" sits under Custom Deployment (enterprise), not under the API (R19-S104). Region: "hosted and operated in the United States" (Privacy Policy; R19-S106). Your data goes to Liquid AI (Jev's route data terms do not apply). Routes: Liquid direct; Vercel AI Gateway liquid/d1-ultra, listing fields zdr: "none" and no_training: "none" quoted as listed (R19-S28); OpenRouter model page liquid/d1-ultra (R19-S107), not in OpenRouter's /api/v1/models list, which also has no other decision-model ID (19 Oct 2026, 05:22 UTC; R19-S24). Cloudflare's catalog: not checked | Vendor claims only; no graded comparison. Reported Liquid AI, which sells a competing hosted decision model, says "d1-ultra outperforms Jev 1.14 across all tasks" over 112 evaluations it ran itself (launch post; R19-S100). Not compared with Jev by this pool | 7 Nov, 15:36 UTC (pricing page and docs: price and weights line only, unchanged; R27-S30). 31 Oct, 05:11 UTC (launch post, docs and pricing page: status, price, image limit and credit unchanged; billing sentence reworded; R25-S211 to R25-S213). 19 Oct, 05:22–05:23 UTC (launch post, docs, pricing page, Terms, Privacy Policy, OpenRouter and Vercel listings) |
| Mouse-1 (Mouse) (preview; launched 13 Oct 2026) · API docs · pricing page | Hosted API from another provider: the whole hosted service. A different provider and model (mouse-preview-2026-10-13) with its own POST /v1/decisions and a /v1/systemone alias in Jev's request shape. Not a product built on Jev | Not applicable (hosted, proprietary). Weights not published; base model not named (19 Oct 2026, 05:24 UTC; R19-S123) | Documented by Mouse. Status and request: Preview: the aliases mouse-1 and mouse resolve to mouse-preview-2026-10-13, Status "Preview" (models and versioning docs, no date shown, read 19 Oct 2026, 05:24 UTC; R19-S131; docs, checked 7 Nov 2026, 15:36 UTC, unchanged, R27-S31). Request: POST https://api.mouse.dev/v1/decisions with a Bearer key; "POST /v1/systemone is served as an identical alias for TypeSafe System One SDKs" (R19-S122); not tested. Input: text and images; "Image input is in beta" (R19-S123). Limits: state 16,384 tokens, 128 questions, 8 images, 4 MB body (R19-S129). Errors and limits: 400, 401, 402 insufficient_credits, 413, 422, 429 (four codes, with Retry-After) and 503 engine_timeout; standard key 10 requests per second, 1,000,000 input tokens per minute (R19-S128, R19-S129). Price as published: "Input: $0.10 / 1M tokens Output: $0.00 (free)" (pricing page, no date shown, read 19 Oct 2026, 05:24 UTC; R19-S121; pricing page, checked 7 Nov 2026, 15:36 UTC, unchanged, R27-S31). Access: prepaid credits; "$5 in credits, no card" (launch post; R19-S120); "Verified accounts receive $5.00 in free starter credits at signup." (R19-S130). Your data: training: "We do not use API request content to train or fine-tune models." (Terms, Effective date October 14, 2026; R19-S124). Retention: "We process that content in memory to answer the request, then discard it. We do not write it to logs or databases"; the console activity log for metadata is "deleted after 30 days" (Privacy Policy, Effective date October 14, 2026; R19-S125). Region: "served from the United States (US-East)" (R19-S125). Status (Preview), price ($0.10 per 1M input, output free), $5.00 starter credits, Terms and Privacy Policy (both effective October 14, 2026; same training, discard, 30-day and US-East wording) and the 16,384-token and 128-question limits unchanged at 31 Oct 2026, 05:13 UTC (R25-S215 to R25-S220). Your data goes to Mouse (Jev's route data terms do not apply). Routes: Mouse direct only; not on the OpenRouter or Vercel AI Gateway model lists at 19 Oct 2026, 05:22 UTC (R19-S24, R19-S28) | Vendor claims only; no graded comparison. Reported Mouse, which sells a competing model and an agent app, reports 89.1% average decision accuracy for Mouse-1 vs 79.1% for Jev on "unfamiliar tools and repositories", and a decision flip rate under a planted instruction of 0% vs 65.4% for Jev, from its own tests (launch post; R19-S120). Not compared with Jev by this pool | 7 Nov, 15:36 UTC (pricing page and docs: price and aliases only, unchanged; R27-S31). 31 Oct, 05:13 UTC (versions, pricing and limits docs, pricing page, Terms, Privacy Policy: unchanged; R25-S215 to R25-S220). 19 Oct, 05:24 UTC (launch post, docs, pricing page, Terms, Privacy Policy, OpenRouter and Vercel lists) |
| Eevo / Eevo-1 (Upright) (Upright Oy, Finland; announced Oct 2026) · Eevo docs · Terms of Use · Privacy Statement | Hosted API from another provider: the whole hosted service. A different provider and model (eevo-1, now Eevo-1.2) with its own native decision API (options, stakeholders, consequences → expected value per stakeholder) plus a SystemOne-compatible POST /api/v1/systemone. Not a product built on Jev | Not applicable (hosted, proprietary). Weights and base model: not found on the announcement or docs (19 Oct 2026, 05:25 UTC; R19-S140, R19-S148) | Documented by Upright. Status and request: "Eevo is available through MCP, API, and the Eevo console" (announcement, checked 31 Oct 2026, 05:13 UTC; R25-S221); no preview label seen. Earlier wording: "Eevo-1 is available today through the Eevo API, MCP and the Eevo console" (announcement, saved copy of 18 Oct 2026, 05:37 UTC; R19-S140); no preview label found on the announcement or the docs reference pages (19 Oct 2026, 05:25 UTC). Model: "eevo-1 ... (now Eevo-1.2)"; "eevo-1.1, the previous version, is no longer served, and requests that send it are answered by eevo-1.2" (models-pricing docs, checked 31 Oct 2026, 05:13 UTC; R25-S222). Request: the native API in Default mode (many model calls, "typically around 5 seconds") or Direct mode (one call); the SystemOne-compatible POST /api/v1/systemone on eevo.io takes 1–32 typed questions and a state up to 130,000 characters (models-pricing docs, checked 31 Oct 2026, 05:13 UTC; R25-S222); not tested. Change, 31 Oct 2026: state limit, model version and status wording changed between 19 Oct and 31 Oct 2026; the pages show no date. Earlier: a state up to 50,000 characters (R19-S148, R19-S152). Also an MCP server. Text only: no image field found (R19-S152). Errors and limits: 400, 401, 402 insufficient_credit, 409, 429 (Retry-After) and 503 free_capacity_busy; 600 requests per minute per key, 1,200 per organization (R19-S150, R19-S151). Price as published: two prices, side by side as Upright lists them: "$0.15 per decision" in Default mode, and $0.10 per 1M input tokens in Direct mode and the SystemOne-compatible API ("eevo-1 ... (now Eevo-1.2) ... | $0.15 per decision | $0.10 per 1M input tokens"); "Billed per input token; output tokens are free." (models-pricing docs, no date shown, checked 31 Oct 2026, 05:13 UTC; R25-S222, R25-S45). The announcement says "Eevo weighs the trade-offs in seconds, for $0.15 per decision" (quoted as published; R25-S221). Free allowance: "Your first account starts with 5 decisions, each Default mode decision revisable up to 10 times, free" (R25-S222); free Direct-mode decisions are not mentioned. Change, 31 Oct 2026: prices, version and free allowance changed between 19 Oct and 31 Oct 2026; the pages show no date. Earlier (read 19 Oct 2026, 05:25 UTC; R19-S148, R19-S140): "eevo-1 ... $2.50" per 1M input tokens; "Output tokens are free"; Default mode billed "summed over every model call of the request", and "there is no fixed maximum price"; the announcement said "for around $0.20 per decision"; new accounts got 5 free Default-mode and 50 free Direct-mode decisions. Your data: training: "the use of User Content to train future Eevo models is governed by your account's data use setting. This setting is on by default and can be turned off at any time in the Eevo console. While it is on, Upright may use the requests and answers your account stores to train future Eevo models in anonymized form." Content sent with keys set to "Not stored" "is never used to train models" (Terms of Use, no date shown, read 19 Oct 2026, 05:25 UTC; R19-S144; Terms of Use, no date shown, checked 7 Nov 2026, 15:36 UTC, unchanged, R27-S32). Retention is set per API key: "not stored, 30 days, 90 days, 1 year or without time limit" (Privacy Statement, no date shown; R19-S145); "Not stored" "can be chosen once your organization has bought credit; until then, requests are stored" (docs; R19-S149). Default retention for a new key: not found on the retention or authentication docs (19 Oct 2026, 05:25 UTC). DPA: "Organization customers can conclude a data processing agreement with us by contacting us" (R19-S145). Training setting, per-key retention options, the "Not stored" condition and the DPA line unchanged at 31 Oct 2026, 05:13 UTC (R25-S223 to R25-S225). Processing region: not found on the Terms, Privacy Statement or docs pages read (19 Oct 2026, 05:25 UTC). Your data goes to Upright (Jev's route data terms do not apply). Routes: Eevo direct only (API, MCP, console); not on the OpenRouter or Vercel AI Gateway model lists at 19 Oct 2026, 05:22 UTC (R19-S24, R19-S28) | Vendor claims only; no graded comparison. Reported Upright, which sells a competing model, publishes a table against "TypeSafe's jev-latest, run through their API on 6 October 2026", mixing public benchmarks with its own evals (for example PokerBench 87.7 for Eevo-1 vs 72.6 for Jev; CogBench 95.2 vs 96.8) (announcement; R19-S140). Not compared with Jev by this pool | 7 Nov, 15:36 UTC (Terms of Use: training wording only, unchanged; prices not re-compared; R27-S32). 31 Oct, 05:13 UTC (announcement, models-pricing docs: prices, version, free allowance and state limit changed; Terms of Use, Privacy Statement and retention docs unchanged; R25-S221 to R25-S225; pricing table lines read again 05:18 UTC, R25-S45). Errors and authentication pages not re-read (19 Oct). 19 Oct, 05:25 UTC (docs, Terms of Use, Privacy Statement); announcement from the saved copy of 18 Oct, 05:37 UTC |
Drex (Nace.AI) ("Generally available" per its launch post of 24 Sep 2026) · Console docs · pricing docs · Hugging Face nace-ai · nace-ai/drex-decision-models (be8ba7c) | Hosted API from another provider: the whole hosted service. A different provider and model (drex-v2.0, which drex-latest points to) at its own POST /v1/systemone in Jev's request shape; also independent open weights (Drex 2.0 and 1.5) with a local runtime (counted once, under hosted). Drex Max is hosted only. Not a product built on Jev (R20-S102, R20-S105, R20-S123) | Code Apache-2.0 (nace-ai/drex-decision-models at be8ba7c; R20-S128). Weights: "NACE AI OPEN RAIL-M LICENSE (MODIFIED) Version 1.0" (LICENSE at nace-ai/drex-v2.0 3e6d164; byte-identical at nace-ai/drex-v1.5 f80f9d2): a separate commercial licence is required if you (or your employer) had more than "$1,000,000" in gross revenue in the prior 12 months or raised more than "$1,000,000", "except where Your Use is limited to personal use or research purposes" (R20-S122, R20-S124). Base: Qwen/Qwen3.6-35B-A3B (2.0) and XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (1.5) per card metadata Reported; base licences not read (R20-S123, R20-S125). Drex Max: no open weights (R20-S105) | Documented by Nace.AI. Status and request: "Generally available" (launch post, 24 Sep 2026; R20-S100); no preview or waitlist label found in the docs (20 Oct 2026, 05:17 UTC; R20-S105 to R20-S108). Request: POST https://console.nace.ai/v1/systemone with a Bearer key; body model, state (text or JSON) and questions (noul, choice, score); 1–512 questions (Drex Max 16), body at most 4,194,304 bytes, state up to 131,072 tokens (R20-S102, R20-S107); Drex Max also takes up to 8 images (R20-S102). Nace's docs say "Drex is wire-compatible with TypeSafe's Jev" and that TypeSafe's SDK works "without a fork" Reported (vendor statement; not tested; R20-S110). Errors and limits: 401, 402, 422, 429 and 529 (with retry-after), 500; free tier 120 requests per minute, paid 600 (R20-S107, R20-S108). Price as published: "Per 1M input tokens, drex-v2.0 $ 0.06"; drex-v1.5 $0.05; drex-max $0.06; output tokens "not billed" (Console pricing docs, no date shown, read 20 Oct 2026, 05:17 UTC, re-read 05:23 UTC; R20-S106, R20-S51). Change, 30 Oct 2026: On 30 Oct 2026, 05:19 UTC the same pricing docs list drex-max at $0.03 per 1M input tokens (was $0.06 on 20 Oct), a new drex-max-v1.1 at $0.03, drex-v2.0 $0.06 and drex-v1.5 $0.05, and a signup balance of $5.00 (R24-S225). Changed between 20 Oct and 30 Oct 2026; the page shows no date. Same four prices and signup balance at 31 Oct 2026, 05:11 UTC (R25-S214). drex-v2.0 (60 nUSD per input token) and drex-max ($0.03 per 1M input tokens) unchanged at 2 Nov 2026, 10:54 UTC (R26-S34). Pricing docs (drex-v2.0 "$0.00000006 ( 60 nUSD)" per input token, "$ 0.06" per 1M; drex-max-v1.1 $0.03 per 1M) and the product page ("$0.04" per "1,000 decisions") checked 7 Nov 2026, 15:34–15:36 UTC, unchanged (R27-S29). drex-v1.5: "$0.04" per 1,000 decisions on the Drex product page; "$0.05" per 1M input tokens in the pricing docs (both read 21 Oct 2026). The two use different units; the pool does not convert one into the other (R20-S102; 21 Oct: R21-S240, R21-S241, R21-S242). Access: account and key; Nace's docs say new accounts get "$25.00 of signup credit once a person is behind it" (an access condition as Nace states it, not a statement of what Drex costs; R20-S104; read 20 Oct 2026). On 30 Oct 2026, 05:19 UTC Nace's pricing docs show a signup balance of $5.00 (R24-S225); the docs page with the $25.00 wording was not re-read. Your data: training: "No training by default … unless you separately and expressly authorize that use in writing (including through an opt-in control in the Console)" (Service Terms, last updated October 19, 2026, s.5.4(a)); retention: "within 48 hours after processing completes for Paid Accounts and within 7 days after processing completes for Free Tier Accounts" (s.5.7(a)); with zero retention, which an owner or admin can turn on: deleted "within 1 hour after results are delivered" (s.5.7(g)) (R20-S111; re-read 05:23 UTC, R20-S52). Website terms vs Console terms: Documented Nace.AI's own legal pages disagree. Its website Terms and Privacy Policy (both dated 9 Oct 2026) say customer content may be used for training by default, with an opt-out, and kept up to 30 days. Its Console Service Terms and Data Processing Addendum (both dated 19 Oct 2026) say no training unless you opt in, and deletion within 48 hours on paid accounts or 7 days on free ones. Both sets of Terms say they cover the Console and the APIs. Checked 21 Oct 2026 (R21-S232, R21-S233, R21-S237, R21-S236; markdown copies R21-S234, R21-S235). Which document controls is not stated on the pages read; the catalogue does not infer it. The Console Data Processing Addendum describes an optional zero-retention mode: deletion within 1 hour, nothing written to logs or backups. Its sub-processor schedule still says processing locations are "to be confirmed before publication" (21 Oct 2026; R21-S236). The Addendum also carries draft-like lines ("Pending: … to be implemented and verified"), quoted here as text Documented; whether those controls exist is Unverified. The sub-processor list at trust.nace.ai/subprocessors, linked from the website Privacy Policy, returned HTTP 403 and was not read (R21-S239). Region: "the United States and other countries where our service providers operate" (Privacy Notice; R20-S112). Your data goes to Nace.AI (Jev's route data terms do not apply). Routes: Nace direct; OpenRouter pages for nace-ai/drex-v1.5 ($0.04 / $0 per 1M) and nace-ai/drex-v2.0 ($0.06 / $0 per 1M), served by DeepInfra, with OpenRouter's fields training:false and retainsPrompts:false (quoted, not interpreted); neither ID was in OpenRouter's public /api/v1/models list at 20 Oct 2026, 05:16 UTC (R20-S116 to R20-S118) | Vendor claims only; no graded comparison. Reported Nace.AI, which sells a competing hosted decision model, reports Drex 2.0 at "61.25 vs 57.96" against Jev 1.13.0 on Decision Index 0.3.1 public, and Drex Max "6.9 points over Jev 1.14", from its own figures (product page and model card; R20-S102, R20-S123). The phrase "wins the Decision Index by a rounding error" is a daily.dev headline (third-party aggregator); it was not found on the Nace.AI pages read on 20 Oct 2026 (R20-S131). Not compared with Jev by this pool | 20 Oct, 05:16–05:19 UTC (launch post, product page, docs, pricing docs, Service Terms, Privacy Notice, Hugging Face cards and LICENSE files, GitHub LICENSE, OpenRouter pages); price and data terms re-read 05:23 UTC; 21 Oct, 05:12–05:15 UTC: website Terms and Privacy Policy, Console Service Terms, Privacy Notice and Data Processing Addendum, product page and pricing docs (price units); trust.nace.ai/subprocessors HTTP 403, not read; 30 Oct, 05:19 UTC: pricing docs only (prices and signup balance changed); 31 Oct, 05:11 UTC: pricing docs only (unchanged; R25-S214); 2 Nov, 10:54 UTC: pricing docs, drex-v2.0 and drex-max prices only (unchanged; R26-S34); 7 Nov, 15:34–15:36 UTC: pricing docs and product page, prices only (unchanged; R27-S29) |
| Base Labs (Baseten): mammoth-1, elephant-1, joey-1 (Base Labs, Baseten's research lab: labs.baseten.co; R27-S207, R27-S209) · pricing page · Model APIs docs As the library pages name them: mammoth-1 decision-models/baseten-mammoth-27b-latest (27B); elephant-1 decision-models/baseten-elephant-9b-latest (9B); joey-1 decision-models/baseten-joey-2b-latest (2B) (R27-S200 to R27-S202, read 7 Nov 2026, 15:30 UTC)Rank claims checked against the Jev Decision Index on 24 Oct 2026 | Hosted API from another provider: the whole hosted service. A different model; your data goes to Baseten (Base Labs' parent company) under Baseten's terms. Served on Baseten's Model APIs at https://inference.baseten.co/v1/systemone with an Api-Key header (R27-S200 to R27-S202). Not a product built on Jev | Hosted service: Baseten's terms (below). Weights published by Base Labs on Hugging Face (baselabs/mammoth-1, elephant-1, joey-1; licence MIT on the cards and library pages; base models Qwen/Qwen3.5-27B, Qwen/Qwen3.5-9B and Qwen/Qwen3.5-2B per card metadata). Not counted under model weights (7 Nov 2026, 15:30 UTC; R27-S200 to R27-S206) Documented | Documented by Baseten. Status: live on Baseten Model APIs: "Start using mammoth-1 instantly with per-token pricing" (same wording for elephant-1); "joey-1 is live on Baseten Model APIs with zero data retention by default"; no preview, beta or waitlist label found on the library pages (7 Nov 2026, 15:30 UTC; R27-S200 to R27-S202, R27-S209). The 21 Oct launch post's "free to try for a week on Baseten Model APIs" refers to a launch week that has ended (R27-S208). Signup: the "Get started" link leads to a self-serve sign-up page (observed, not used). Price as published: "Token pricing ($ per 1M tokens)", Input / Cache input / Output: mammoth-1 $0.10 / — / —; elephant-1 $0.035 / — / —; joey-1 $0.01 / — / — (pricing page, read 7 Nov 2026, 15:30 UTC, re-read 15:39 UTC; R27-S203, R27-S42); the library pages show the same input prices under "Pricing per 1M tokens" (R27-S200 to R27-S202). Same prices as on 30 Oct 2026 (benchmarks line; R24-S223). TypeSafe's published Jev price is $0.042 per million input tokens, output free (cost per decision). Your data: zero data retention: "For synchronous Model API requests, Baseten doesn’t store inputs or outputs under its zero data retention policy. For queued execution, use async inference with a dedicated deployment and review its input-retention behavior." (Model APIs docs; R27-S212). Training: "6.3. No Training on Customer Content. Baseten will not use Customer Content, including Customer Models, Model Outputs, or inference logs, to train, fine-tune, or otherwise develop machine learning or artificial intelligence models." Telemetry: "Baseten may collect technical and operational telemetry relating to the use and performance of the Services […] and may use Usage Information to operate, maintain, secure, and improve the Services." (Terms and Conditions, s.6.3, s.6.4; R27-S210). Retention: "Baseten will retain Customer Content in accordance with a customer's instructions and as required by applicable law." (Privacy Policy, s.7; R27-S211). All read 7 Nov 2026, 15:30 UTC. Your data goes to Baseten (Jev's route data terms do not apply). Routes: Baseten (Base Labs), first-party. Also on OpenRouter as baseten/mammoth-1, baseten/elephant-1 and baseten/joey-1 (provider BaseTen, 16,384-token context, the same per-token prices; the :free variants list no endpoint), reached through OpenRouter's per-model endpoints pages; not in OpenRouter's public /api/v1/models list (460 entries, 7 Nov 2026, 15:29 UTC). OpenRouter's own data terms for these endpoints were not read (R27-S213 to R27-S216, R27-S218; re-read 15:38 UTC, R27-S42) Reported | Vendor claims only; no graded comparison. Reported Base Labs, which sells competing decision models through Baseten, cites the Jev Decision Index 0.4: mammoth-1 "#1 of 208 models on the Jev Decision Index 0.4: 71.24, ahead of Perplexity Decider v1.2 (69.42), Jev 1.14 (+9.0) and Cloudflare Clef (+15.5)"; elephant-1 "Best model under 15B on the Jev Decision Index 0.4: 62.65, tied #8 of 208 models"; joey-1 "Best model under 4B on the Jev Decision Index 0.4 public split: 63.30, #37 of 230 models" (Hugging Face cards, 7 Nov 2026, 15:30 UTC; R27-S204 to R27-S206). The cards cite 208 and 230 models, which suggests different board dates. Not matched against the board on 7 Nov 2026: its data returned HTTP 500 at 15:31 UTC (R27-S220). The benchmarks page checked the mammoth-1 claim on 24 Oct 2026: it matches the index's version 0.4 data of 20 Oct 2026 and not its version 0.4.1 data of 22 Oct 2026 (operator, versions and commits there). Not compared with Jev by this pool | 7 Nov, 15:29–15:31 UTC (library pages, pricing page, Hugging Face cards, Base Labs site, launch posts, Terms, Privacy Policy, Model APIs docs, OpenRouter endpoints and public list; board data HTTP 500); pricing page and OpenRouter endpoints re-read 15:38–15:39 UTC (R27-S42) |
| GLiClass | Classifier or LLM workflow | Not recorded | Local or self-hosted text classifier; not Jev's full typed-question contract | None | 22 Sep, not rechecked |
| LLM with structured output | Classifier or LLM workflow | Provider terms | Generative schema-constrained output; no Jev-style probabilities | None; schema-valid is not correct. Published Jev-vs-LLM results: head-to-head table | 22 Sep, not rechecked |
Documented Licences, runtimes and API shapes as stated in each repository or model card on the check date. "Not recorded" means the pool did not record it; it does not mean there is no licence. Reported All accuracy, speed and calibration figures are the authors' or operators' own.
Licence traps: do not carry a licence or a benchmark from one artifact to another. he-jev/laya is a copy of the Hugging Face model repository convaiinnovations/laya at commit c5d78730 (19 Sep), not of the linked code repository; who runs it is still not established (checked 1 Oct 2026). The separate zhangcy122/OpenJev remains an unchecked lead. Three more rows checked on 22 September (SemIf/OpenJev, jevlike and mini-jev) are kept in the 22 September substitution report.
Seen, not yet checked
Seen in the 8–10 Oct scan, not yet checked (listed for the next run; descriptions only; not rows; R12-S02, R12-S13, R12-S14, R12-S17): THX-01 (Hugging Face doofz/THX-01, PyPI thx01; "beats Jev", Reported by its author), Decision Studio (BudEcosystem; a desktop runtime for Jev-style models), Jevfree (EmbeddingGemma 2), CanadaOrNaw/dex, javimosch/jev-nano, fstandhartinger/decisionmodels-local, pedroliveiraAI/LLMtoSystemOne, gtokman/iKev, RetroUpgrade-YouTube/Strata-JEV, JaakkoLipp/luna-middleware-decisions (turns OpenAI's model into a Jev-like endpoint: an interface over another provider's hosted API) and agarwalastitwa027-svg/mini-jev-multimodal. Unverified
Seen in the 11 Oct scan, not checked (names only; not rows; to be followed to a primary source before any row; R13-S22–S27): standardthinkingai/standard-one-sglang (a runtime fork), Louiskq/jev-class-9b (open weights), ronitgupta138/local-jev, nirvam/log-pi (GGUF behind a System One-compatible API, as described), miniproaissa/laya-browser-agent and "Cribl Decision 1.0". Unverified
Seen 13 Oct 2026, not checked (not rows; their primary sources have not been read): Cygnet (named in B07 with a score of 79.0) and GLiNER2.5-Decide (named in B29 as "an open decision model"): no primary source read (comparator review, 13 Oct 2026; R15-S56). Unverified SubQ Snap, listed here on 13 Oct, has been a row since 14 Oct 2026.
Seen 19 Oct 2026, not checked further (not a row): Tazo Labs (Kodimo, Inc. DBA Tazo Labs) describes an audio decision model: audio plus a question and fixed answers in, a yes/no, pick-one or score answer with a probability out (blog, 4 Oct 2026). No price or request shape found in that post (saved copy, 18 Oct 2026, 05:29 UTC; R19-S160). API access is by early-access request. Reported
These appeared during the 27–28 September scan. They are leads only: no licence, parity or safety claim is made beyond what each project states. Unverified
- Jeva.cpp: a llama.cpp fork with a "JEV-compatible decision API" (repository states MIT).
- Gevva0: a local decision gateway on Gemma (repository states Apache-2.0).
- Rene-1 31B: claims "+9 over Jev" on a "Decision Index"; a benchmark claim only (the card marks its score
verified: false; graded D, B18). The card says Apache-2.0; the base model is not stated in the card metadata (checked 30 Sep 2026). - Jeff (firelex): repository created 28 Sep, v1.1 on 29 Sep; code MIT, weights licence not checked. The README says "On benchmarks they approach, and sometimes beat, Jev" (graded D, B20). Not the same project as the Gestalt-Lab or logan-markewich Jeff rows above.
- llama-system1: an npm shim exposing
/v1/systemonefor llama.cpp. - OllamaMQ v0.5.0: its release notes say it "now supports decision-model backends ("System One" models)"; it is a queue in front of such servers, not a model (checked 10 Oct 2026, 05:14 UTC; R12-S136).
- kev-onnx: a CPU-only ONNX server for Kev with a
/v1/systemoneAPI. - AnyJev v0.2.0 was released on 28 Sep. Whether it is the same project as the "JevAny" row above was not verified, so they are kept separate.
Two videos compare Laya and Kev without running Jev: Zahiruddin Tavargere's re-run on Laya only, and RUNTIME's "Kev vs Laya" (its description: "30 frozen synthetic cases … Kev matched our policy 26/30 times and Laya English 12/30"). These are author-run comparisons, not Jev demos, and are therefore not listed on the videos page. Reported The graded tests that ran Jev with Laya or Kev are collected on Jev vs Laya vs Kev.
Jev vs Kev: one operator-reported comparison
Opper's 25 September article, with benchmark code, a 362-item manifest and results, reports one run on 160 arXiv labels, 120 Stack Exchange sites and 82 GitHub issue labels. Reported It is row B11 (grade B; flags COI, small-n) on the benchmarks ledger.
| Measure (Opper's own run) | What Opper reports | How to read it |
|---|---|---|
| Accuracy | Both models roughly 94–98% on all three tasks | Relatively easy tasks; small gaps do not show general superiority. |
| Calibration error | Lower for Jev on all three tasks | One suite, one run. |
| Counted input tokens | About 257 more per Jev request | Route- and workload-specific; the 12× example is one short request, not a general multiplier. Used on Cost per decision. |
| Concurrent median latency | 220 ms (Jev) vs 275 ms (Kev) | Depends on route, region and load. |
Opper sells route access to Jev and hosts one of the compared models, so it is not a neutral party. This pool did not rerun the suite.
What was not verified
- Same-workload, same-hardware parity test of any alternative against Jev: none run by the pool as of 14 Oct 2026. The graded comparisons are B11 (Opper, grade B), five with OpenAI's Decisions API (B38 and B39, grade A since 11 Oct 2026; B41, grade A since 12 Oct 2026; B42, grade B since 13 Oct 2026; B44, grade A since 14 Oct 2026), two with Microsoft-Decision-1 (B41, B44) and one with SubQ Snap (B43, grade B since 14 Oct 2026). On 24 Oct 2026: six with OpenAI's Decisions API (B45, grade A, added) and two with SubQ Snap (B46, grade A, added); Microsoft-Decision-1 unchanged at two. B32 (grade A) uses Cloudflare's Clef and Clef-flash as comparators.
- Base-model licences and some checkpoint and adapter licences (JevAny adapters, vjev-vision): not established; not rechecked since 1 Oct 2026.
- "Local" is not a privacy guarantee: check downloads, network calls, logging and retention for each project.
- Rows keep their own check dates (22 September–12 October). On 1 October only Kev and Laya were rechecked; on 10 October only Strands Decider was rechecked and Jiwo 0.8B, TOD and the Decisions API were added. On 11 October Microsoft-Decision-1 was added and the Decisions API evidence cell was rebuilt from B38 and B39; no other row was rechecked. On 12 October the Microsoft-Decision-1 row was rebuilt, the Decisions API row was rechecked (data controls and B41 added) and the two Cloudflare Clef rows were added; no other row was rechecked.
- Not read on 10 Oct: the Enterprise page's on-premises answer; the second page of Ollama's search; the Strands LICENSE.md text (metadata only), strandsagents.com and the Gemma 4 base licences of the Strands variants; TOD's base-card licence and its hosted API shape; OpenAI's pricing page (only the guide was read).
- Not read on 11 Oct: Microsoft-Decision-1's data terms, licence terms and request shape; OpenRouter's terms for it; the
kartikanand73/decision-api-benchmarkrepository beyond its README. (Request shape, data page and the repository were read on 12 Oct.) - Not read on 12 Oct: Microsoft's Product Terms, a Decision-1 retention period in days, the catalog's License tab, the regions offering DataZoneStandard, the Jev benchmarks added to Microsoft's post; OpenRouter's terms; for Cloudflare Clef: the Qwen base-model licences, Cloudflare's Self-Serve Subscription Agreement and Privacy Policy, timeout and error behaviour, the Hugging Face model cards' own text, Cloudflare's benchmark materials, and whether Clef is offered on other routes.
- All authors' accuracy figures, including those of Jiwo 0.8B, TOD and Strands Decider, are their own. Of the Decisions API comparisons, B38 and B39 (11 Oct 2026) and B41 (12 Oct 2026) are graded; classifier-bakeoff is a lead, not graded (metric code not opened).
- Strands Decider's AWS link: no aws.amazon.com source and no AWS-organisation GitHub source was found (10 Oct 2026, 05:13 UTC); the
strands-labsGitHub organisation is unverified and has no description. - Not checked on 19 Oct (d1-ultra, Mouse-1, Eevo): no model was called, so the vendors' statements that TypeSafe's SDKs work against their endpoints are untested; Liquid's error codes and rate limits were not found; Cloudflare's catalog was not checked for d1-ultra; Eevo's processing region was not found, its legal pages show no date, and its status sentence is from the 18 Oct saved announcement; Vercel's
zdrandno_traininglisting fields are quoted, not interpreted. - Not checked on 20 Oct (Drex): no model was called, so Nace's statement that Drex is wire-compatible with Jev is untested. Opened but not read: Nace's www Terms and Privacy pages and the console DPA. Not read: DeepInfra's terms (the OpenRouter route's host), Drex Max's OpenRouter page and launch-post body, the base-model licences, and whether the code files in the Hugging Face repositories fall under the RAIL-M licence or Apache-2.0. OpenRouter's
trainingandretainsPromptsfields are quoted, not interpreted. - Not read on 2 Nov: the contents of Laya's new release v0.4.21 (GitHub; PyPI 0.4.21 published 1 Nov 2026, 21:16 UTC; R26-S33, R26-S118) and of Kev's kev-1.0 release.
- Not read on 31 Oct: the contents of Kev's kev-1.0 release and of Laya's v0.4.18 release; the terms of Makora and DeepInfra (OpenRouter hosts for Clef and Clef-flash); the Decision-1 catalog text, context and licence (rendered in the browser); OpenAI's regional data-controls table and error-codes guide; Eevo's errors and authentication pages.
- The pool ran nothing: no package installed or run, no model downloaded, no benchmark run, no Jev call.
- Ollaya's per-model licences, the origin of its benchmark and Kev-9B's base licence were not checked. Laya's and Kev's System One compatibility is stated by their authors, not tested by the pool.
Sources for the changes since 10 Oct
- R12-S100–S110: typesafe.ai (
/modelsHTTP 404), docsllms.txt,models.md,introduction.md, the Enterprise page, TypeSafe's GitHub organisation listing, Hugging Face author and name searches, Ollama search for "jev"; 10 Oct 2026, 05:11–05:13 UTC. R12-S21 (docs.typesafe.ai/models.md), R12-S36 (GitHub organisation, 12 repositories), R12-S58 (typesafe.ai/enterprise). - R12-S111–S117: Hugging Face API and READMEs for
autotrust/JEV-27B-VL(ba3f0d5), the three MLX conversions,shgao/rsi-jev-v6.1-vl-4b(7f93b19),chenz53/Jev-alpha-26B-A4B(97e483c). - R12-S118–S125:
strands-labs/strands-deciderrepository, commits (HEAD0ac22a9), tags and releases, README;StrandsAgents/strands-decider-2B-qwen3.5-v1-2610(216d1cc);amazon/strands-decider-2B-qwen3.5-v1-2610(974685f, created 9 Oct, 22:55 UTC); Amazon's Hugging Face organisation overview (verified);Qwen/Qwen3.5-2B-Base(b1485b2). - R12-S126–S128:
eljiwo/jiwo-0.8b(52d217f) card and README; GitHubjiwidi/jiwoat8716aba;Qwen/Qwen3.5-0.8Bmetadata. - R12-S129–S133: Hacker News item 49981275; parseclab.ai/tod;
parsecai/tod(07d34d5); GitHubdaseinlabs/todat642834e. - R12-S134 (10 Oct, 05:12 UTC) and R11-S300 (8 Oct, 05:31 UTC): OpenAI Decisions API guide.
- R12-S135: Ollama 0.35 blog. R12-S136: OllamaMQ v0.5.0 releases.
- R12-S11: Google, Bing and DuckDuckGo autocomplete, 10 Oct 2026, 05:17 UTC (Bing and DuckDuckGo returned identical lists, so they may share a backend). R12-S02, R12-S04, R12-S13, R12-S14, R12-S17: the 8–10 Oct scan (GitHub, Hugging Face, Hacker News).
- 11 Oct: R13-S32 (Microsoft's Microsoft-Decision-1 post, 11 Oct 2026, 05:22 UTC; price quoted from it); R13-S33 (Hacker News item 50024913, 9 Oct, 18:38 UTC; read 05:21 UTC). OpenRouter
/modelslist fetched 05:17 UTC in the 11 Oct scan (recorded in the scan note with R13-S32). Decisions API comparisons: graded rows B38 (R13-S50, R13-S52, R13-S53) and B39 (R13-S55, R13-S57–S59), 05:23 UTC. Report: 11 Oct research report. - 12 Oct, Microsoft-Decision-1 and OpenAI: R14-S207 (Microsoft's post, rechecked 05:13 UTC), R14-S208 (Microsoft Learn how-to, Last updated on 2026-10-09), R14-S209 (Foundry catalog), R14-S210 (models sold by Azure), R14-S211–S214 (OpenRouter
/models, endpoints, model page, providers), R14-S215 (Data, privacy, and security for Foundry Models sold by Azure, Last updated on 2026-05-19); R14-S217 (OpenAI pricing page), R14-S218 (OpenAI data controls), R14-S219 (Decisions guide); 05:13–05:15 UTC. - 12 Oct, B41: kartikanand73/decision-api-benchmark at
1ffa4d5:README.md(L205–241, L248, L295),compare/metrics.py(L61–83, L264–296),runs/compare-20261010-222249Z/comparison.md(L26, L29), run-2report.md(L37),data/cases.jsonlandmeta.json; 05:12 UTC (R14-S200–S206; spot-checked against the raw files, R14-S51). - 12 Oct, Cloudflare Clef: R14-S400–S402 (Workers AI model pages for Clef, Clef-flash and Clef-omni), R14-S403 (Workers AI pricing, 05:33:51 UTC), R14-S404 (Workers AI data usage), R14-S405 (Cloudflare changelog, 9 Oct 2026), R14-S406 (Hugging Face API and LICENSE files for
Cloudflare/clef,clef-flash,clef-omni, 05:34 UTC); R14-S106 (ra2webclient.pyat0e4e8cf, L27–33, default modelclef-flash). Report: 12 Oct research report. - 12 Oct, Named like Jev: R14-S31 (
autotrust/JEV-27B-VL, Hugging Face API, 05:16 UTC), R14-S06 (TypeSafe models page), R14-S07 and R14-S08 (docsllms.txt, typesafe.ai). Research note:small-items.md§2. - 13 Oct: R15-S44 (Foundry catalog, lifecycle and licence text, 05:20 UTC), R15-S200 to R15-S204 (Product Terms, model-specific terms, data page, abuse monitoring), R15-S206 (Foundry quotas and limits), R15-S209 (Microsoft's post, charts decoded), R15-S212 (Vercel AI Gateway model list), R15-S213, R15-S214, R15-S216, R15-S217 (OpenAI Decisions guide, API reference, error-codes guide, data controls), R15-S49 (OpenRouter Decisions endpoint), R15-S222, R15-S223 (Workers AI limits and errors), R15-S226 (OpenRouter Clef endpoints), R15-S219 to R15-S221 (Qwen base cards and LICENSE files), R15-S33 (SubQ Snap post), R15-S53 (B42 re-pin), R15-S56 (comparator review); 05:18–05:24 UTC. B42 files: R14-S300 to R14-S305. Report: 13 Oct research report.
- 14 Oct: R16-S217 to R16-S225 (SubQ Snap docs index, decisions API page, data-retention page, models page, docs home, Terms, privacy policy, launch post and home page, 16:47–16:49 UTC), R16-S226 and R16-S227 (OpenRouter and Vercel AI Gateway model lists, 16:47 UTC), R16-S66 (B43 files at
b754c44, 16:52 UTC), R16-S67 (B44 files at285c803, 16:55 UTC). - 16 Oct: R17-S05 and R17-S06 (TypeSafe models and images pages, 05:21–05:22 UTC), R17-S33 (Foundry catalog, Decision-1 lifecycle, 05:21 UTC), R17-S62 (Upright's Eevo page and its Hacker News item, 05:31 UTC). Report: 16 Oct research report.
- 18 Oct: R18-S50 (Liquid AI d1-ultra post, dated 15 Oct 2026; read 05:29 UTC), R18-S51 (Mouse-1 post, dated 13 Oct 2026; read 05:29 UTC). Report: 18 Oct research report.
- 19 Oct: Liquid AI d1-ultra: R19-S100 (launch post), R19-S101 to R19-S103 and R19-S111 (Liquid docs), R19-S104 (pricing page), R19-S105 (Terms of Service), R19-S106 (Privacy Policy), R19-S107 (OpenRouter model page), R19-S108 (Vercel model page); 05:22–05:23 UTC. Mouse-1: R19-S120 (launch post), R19-S121 (pricing page), R19-S122 to R19-S131 (API page, docs, Terms, Privacy Policy); 05:24 UTC. Eevo: R19-S140 (announcement, saved copy of 18 Oct 2026, 05:37 UTC), R19-S141 to R19-S143 (saved Upright pages), R19-S144 (Terms of Use), R19-S145 (Privacy Statement), R19-S146 to R19-S152 (Eevo docs); 05:25 UTC. Route lists: R19-S24 (OpenRouter
/api/v1/models), R19-S28 (Vercel AI Gateway/v1/models), 05:22–05:24 UTC. Tazo Labs: R19-S160 (blog post, saved copy of 18 Oct 2026, 05:29 UTC). Report: 19 Oct research report. - 20 Oct: Drex (Nace.AI): R20-S100 (launch post, 24 Sep 2026), R20-S101 to R20-S103 (Nace.AI home and Drex product page), R20-S104 to R20-S110 (Console docs: overview, models, pricing, limits, errors, API reference, migration guide), R20-S111 and R20-S112 (Service Terms and Privacy Notice, last updated 19 Oct 2026), R20-S116 to R20-S118 (OpenRouter model pages and
/api/v1/models), R20-S119 to R20-S125 (Hugging Face metadata, LICENSE files and model cards at3e6d164andf80f9d2), R20-S127 and R20-S128 (GitHubnace-ai/drex-decision-modelsatbe8ba7cand its LICENSE), R20-S131 (daily.dev); 05:16–05:19 UTC; pricing docs and Service Terms re-read 05:23 UTC (R20-S51, R20-S52). Report: 20 Oct research report. - 23 Oct: OpenAI Decisions API status: R22-S32 (Decisions API guide, 05:23 UTC), R22-S47 (OpenAI API changelog, entry "Oct 22", 05:25 UTC).
- 24 Oct: OpenAI Decisions API caching prices: R23-S32 (Decisions API guide, 05:15 UTC), R23-S45 (OpenAI pricing page, 05:15 UTC), R23-S46 (OpenAI API changelog, 05:15 UTC). B45: gbolask24/decision-bench at
e5b7202: README,dbench/score.py(L23–28, L72–116),dbench/run.py(L21–24),results/summary.json,results/data-pack.md,.gitignore,DEVIATIONS.md; 05:15–05:21 UTC (R23-S200 to R23-S211). B46: jcarlosrodicio/subq-vs-jev atc037905: README,bench/analyze.py(L30–40, L115–118, L131–134),bench/client.py(L12–13),site/data/results.json(L100–104); 05:16–05:21 UTC (R23-S217 to R23-S223). - 30 Oct: OpenAI Decisions API: R24-S200 (Decisions API guide), R24-S201 (OpenAI pricing page), R24-S202 (OpenAI API changelog), R24-S203 (OpenAI "Your data" page), 05:17 UTC. Microsoft-Decision-1: R24-S208 (Microsoft post), R24-S209 (OpenRouter endpoints), 05:19 UTC. SubQ Snap: R24-S212 (docs models page, modified 28 Oct 2026), R24-S213 (docs
llms.txt), R24-S214 (launch post), 05:17 UTC. Cloudflare Clef: R24-S216 to R24-S218 (Workers AI model pages for Clef, Clef-flash and Clef-omni), R24-S219 (Workers AI pricing, Last updated Oct 20, 2026), 05:17 UTC. - 31 Oct: OpenAI Decisions API: R25-S206 and R25-S207 (Decisions API guide, markdown and HTML), R25-S208 and R25-S209 (API reference, markdown and HTML), R25-S210 (OpenAI API changelog), 05:11–05:12 UTC; R25-S242 (OpenAI "Your data" page), 05:17 UTC. Microsoft-Decision-1: R25-S200 to R25-S205 (Product Terms MCA and EA/EAS, model-specific terms, data page, abuse-monitoring page, Decision-1 how-to), 05:11 UTC; R25-S244 and R25-S46 (Foundry quotas and limits), 05:20 and 05:22 UTC; R25-S245 (how-to, errors and answer types), R25-S246 (OpenRouter endpoints), 05:20 UTC. Cloudflare Clef: R25-S239 and R25-S45 (Workers AI data usage, Last updated Oct 16, 2026), 05:14 and 05:18 UTC; R25-S240 and R25-S241 (Workers AI limits and errors pages), 05:14 UTC; R25-S48 (model pages for Clef, Clef-flash and Clef-omni), 05:27 UTC. OpenRouter endpoints for the Decisions API, Decision-1, Clef, Clef-flash and Clef-omni: R25-S42, 05:19 UTC; Vercel AI Gateway model list: R25-S43, 05:14 UTC. SubQ Snap: R25-S47 (docs models page and launch post), 05:27 UTC. Liquid AI d1-ultra: R25-S211 (pricing page), R25-S212 (docs), R25-S213 (launch post), 05:11 UTC. Drex: R25-S214 (pricing docs), 05:11 UTC. Mouse-1: R25-S215 to R25-S220 (pricing page, versions, Terms, Privacy Policy, limits and pricing docs), 05:13 UTC. Eevo: R25-S221 (announcement), R25-S222 (models-pricing docs), R25-S223 to R25-S225 (Terms of Use, Privacy Statement, retention docs), 05:13 UTC. Kev and Laya: R25-S262 (Kev latest release, tag kev-1.0, 1 Oct 2026) and R25-S263 (Laya latest release, tag v0.4.18, 30 Oct 2026), 05:20 UTC.
- 2 Nov: Laya: R26-S33 (latest GitHub release, 10:54 UTC), R26-S118 (PyPI
laya, 10:48 UTC). Microsoft-Decision-1: R26-S41 (Foundry catalog, 11:00 UTC; catalog text present in the page as fetched), R26-S35 (OpenRouter endpoints, 10:54 UTC). Drex: R26-S34 (pricing docs, 10:54 UTC). Vercel AI Gateway model list: R26-S27, 10:54 UTC. Decisions API, Decision-1 request and data pages, SubQ Snap and Kev: checked 2 Nov 2026 against the 31 Oct values (sources as listed for 31 Oct). - Earlier rows: see the reports linked under Related pages, including the 10 Oct research report.