Hosted decision APIs compared: what changed by 31 October 2026
The comparison of Jev with OpenAI's Decisions API, Microsoft-Decision-1 and Cloudflare's Clef had 40 cells whose latest check date was before 17 October 2026. On 31 October 2026 each was read again at its source where the source could be read. This report lists what changed, what was re-read and found unchanged, and what was not re-read, with the source and time for each. It also records the hosted rows on Alternatives, a new framework package read in source, and the day's access check.
Microsoft now lists Decision-1 limits by subscription tier from 2,160 requests and 2.16M tokens per minute (Tier 1) to 15,000 requests and 15M tokens per minute (Tier 6). On 13 Oct 2026 the same page listed 60 to 900 requests per minute and no token limit.
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 is stated. On OpenRouter, Clef-flash is now also served by Makora and DeepInfra; on a non-Cloudflare host your data goes to that host, not to Cloudflare.
OpenAI's Decisions API guide now says it "evaluates text, images, and JSON". A per-request image limit was not found in the guide, API reference or changelog on 30 or 31 Oct 2026; the 13 Oct 2026 reading showed up to 128 images per request. Upright's Eevo now lists $0.10 per 1M input tokens in Direct mode and its SystemOne-compatible API, and $0.15 per decision in Default mode.
Jev's own column did not change: price, limits, data terms, errors and routes read the same as when last checked.
The 40 comparison cells dated before 17 October: what the 31 October re-read found
- Jev 12 cells12 unchanged
- OpenAI Decisions API 8 cells6 unchanged · 2 changed (inputs, regions)
- Microsoft-Decision-1 10 cells5 unchanged · 1 changed (rate limits) · 2 in part (limits, terms) · 2 not re-read (inputs, graded tests)
- Cloudflare Clef 10 cells6 unchanged · 3 changed (retention, training, routes) · 1 first read: no region wording found
Counts: 29 unchanged + 6 changed + 2 re-read in part + 2 not re-read + 1 first read = 40 cells. Bar length follows the number of cells in each column. "In part": for Decision-1's limits, the OpenRouter listing was re-read but the Foundry catalog text was not (the catalog page builds its text in the browser); for its terms, the Product Terms were re-read but the catalog licence text was not. The two cells not re-read keep their earlier dates on the comparison page. The card at the top of the comparison page repeats these cells and follows their dates.
When each change could have happened, 12–31 October 2026 (UTC)
- Catalogue's daily checks
- Decision-1 rate limits
- OpenAI inputs and regions wording
- Clef hosts on OpenRouter
- Cloudflare data-usage page
- Eevo prices
A bar spans the time between the last read that showed the old value and the first read, or the source's own date, that shows the new one. None of these sources gives the time of its change, except Cloudflare's page, which carries its own update date (16 Oct); the dashed line runs from that date to the first read that saw it. Dots are the days the catalogue ran its daily checks; the comparison cells were not re-read on most of those days, which is why the bars are long. The hatched span is 24–30 Oct, when there was no check.
What changed, and since when
| Fact | Before | Now (as published) | Changed between |
|---|---|---|---|
| Microsoft-Decision-1 rate limits | Requests per minute by tier: 60 (Tier 1) to 900 (Tier 6); no tokens-per-minute figure (13 Oct 2026) | Tokens and requests per minute by tier: Tier 1 2.16M / 2,160; 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. "Your deployment's rate limits depend on its configured capacity and allocated quota." (R25-S244, R25-S46) | 13 Oct and 31 Oct 2026; Microsoft's page shows no new update date |
| Clef on Workers AI: data retention and training | Content "may be stored by Cloudflare if you specifically use a storage service"; no retention period; "does not use your Customer Content to (1) train any AI models …" (12 Oct 2026) | "Zero data retention: Workers AI processes Customer Content only for as long as needed to provide the service." Exceptions: storage services, AI Gateway logging or caching. No period in days is stated. "No training: Cloudflare does not use Customer Content processed by Workers AI to train AI models." (R25-S239) | 12 Oct and the page's update date, 16 Oct 2026; first seen 31 Oct |
| Clef hosts on OpenRouter | cloudflare/clef via PrimeIntellect (16,384 context), DeepInfra and Cloudflare; clef-flash via PrimeIntellect and Cloudflare (13 Oct 2026) | cloudflare/clef: DeepInfra, PrimeIntellect, Cloudflare (65,536-token context; DeepInfra lists 65,537); cloudflare/clef-flash: Makora (16,384), DeepInfra, PrimeIntellect and Cloudflare (24,576); per-host input prices as OpenRouter lists them. On a non-Cloudflare host, your data goes to that host (R25-S42) | 13 Oct, 05:21 UTC and 31 Oct 2026, 05:19 UTC |
| OpenAI Decisions API inputs | "evaluates text, images, or both"; up to 128 images per request (13 Oct 2026) | "The Decisions API evaluates text, images, and JSON". A per-request image limit was not found in the guide, API reference or changelog on 30 or 31 Oct 2026; the 13 Oct 2026 reading showed up to 128 images per request (R25-S206 to R25-S210) | 13 Oct and 31 Oct 2026 (not dated by OpenAI) |
| OpenAI Decisions API regions | "Data residency and regional processing are supported in the United States and Europe (EEA + Switzerland)" (13 Oct 2026) | "Data residency is available in all supported API regions. Regional processing is supported in the United States and Europe (EEA + Switzerland)." (R25-S206) | 13 Oct and 31 Oct 2026 |
| Upright Eevo price and version | eevo-1 $2.50 per 1M input tokens; Default mode with no fixed maximum price; announcement "around $0.20 per decision"; 5 free Default and 50 free Direct decisions; state up to 50,000 characters (19 Oct 2026) | $0.10 per 1M input tokens (Direct mode and the SystemOne-compatible API) and "$0.15 per decision" (Default mode); eevo-1 is now Eevo-1.2, and eevo-1.1 "is no longer served"; a first account starts with 5 free decisions; state up to 130,000 characters (R25-S221, R25-S222) | 19 Oct, 05:25 UTC and 31 Oct 2026, 05:13 UTC |
| Kev and Laya (open models) | Rows describe the versions read earlier | New releases not yet read: Kev kev-1.0 (1 Oct 2026) and Laya v0.4.18 (30 Oct 2026). Nothing about their contents is claimed (R25-S262, R25-S263) | Release dates as GitHub lists them |
Re-read on 31 October and unchanged
| Column or row | As read on 31 Oct 2026 | Checked (UTC) |
|---|---|---|
| Jev (12 cells) | $0.042 per million input tokens, output free; 64k per request and 32k per question (262k with the long-context header); "Rate limits are set per organization"; text by default, images behind an opt-in header; 401, 422, 429 and 529 documented; no training on input; zero data retention "for enterprise customers"; hosted in the United States; six routes list the same IDs (R25-S03, R25-S05, R25-S06, R25-S12b, R25-S19 to R25-S30, R25-S49) | 05:11–05:16 |
| OpenAI Decisions API (6 cells) | 2–255 choices and a refusal answer; refusal wording; "/v1/decisions" not used for training, 30 days of abuse monitoring; OpenRouter and Vercel list openai/gpt-6-luna-decisions; no context window or rate limit stated for the Decisions API (R25-S206 to R25-S209, R25-S242, R25-S247 to R25-S249, R25-S42, R25-S43) | 05:11–05:20 |
| Microsoft-Decision-1 (5 cells, 2 in part) | Request path and authentication; output types; the 400–429 error table; prompts "not used to train, retrain, or improve the base models"; Product Terms still say Microsoft "will temporarily store Input and Output Content" for abuse monitoring, with no period in days and Decision-1 not named; OpenRouter lists it on Azure at 32,768 tokens and $0.042 per 1M input (R25-S200 to R25-S205, R25-S245, R25-S246) | 05:11–05:20 |
| Cloudflare Clef (6 cells) | At most 4 images per request; answers keyed by question; env.AI.run request form; the Workers AI limits page (updated 17 Sep) does not name Clef; the errors table is unchanged in the rows quoted; Apache-2.0 weights. No region wording was found on the Clef or data-usage pages (first read) (R25-S238, R25-S240, R25-S241, R25-S250 to R25-S255) | 05:14–05:20 |
| Hosted rows on Alternatives | Liquid d1-ultra at $0.042 per 1M input tokens, output $0.00, up to 8 images, $2 free credit (its billing sentence reworded to "Each question is billed for its text and all supplied images"); Nace.AI Drex prices and $5.00 signup balance; Mouse-1 in preview at $0.10 per 1M input tokens with $5.00 starter credits; SubQ Snap 1.1 at $0.04 per 1M input tokens in SubQ's docs (its launch post still says $0.05) (R25-S211 to R25-S220, R25-S47) | 05:11–05:27 |
Framework packages: one new row, two newer versions
The framework comparison now has 7 packages read in source: 4 are pre-1.0 and none comes from TypeSafe; 5 raise Jev errors to your code, and 2 decide by themselves. Documented
- TanStack AI,
@tanstack/ai-typesafe1.0.0 (TanStack; read at commit5651e07, tagged 22 Oct 2026): raises to your code on an API error, a malformed answer or a missing key; sets no timeout of its own (a 30 s default is documented but not used). Your code decides; the package does not (R25-S227 to R25-S237). - Pydantic AI,
pydantic-ai-slim2.64.0 (31 Oct 2026): failure path unchanged, still raises to your code. New: an error status that a retrying HTTP transport gives up on now raisesModelHTTPError(R25-S38, R25-S39). - Effect,
@effect/ai-typesafe4.1.1 (30 Oct 2026): failure path unchanged, still raisesAiError; no retry or timeout added. New: image and Locate (Box) decisions, thejev-1.14.0model ID; a malformed response is nowInvalidResponseError(R25-S40).
Not part of the failure map on when Jev fails: framework packages leave the decision to your code, so the map's 80 implementations and its counts are unchanged.
Access on 31 October
No change for a new user: signups are open, new accounts get no free credit, the API is operational, jev-latest still gives jev-1.14.0, and jev-1.13.0 still has no deprecation date (checked 31 Oct 2026, 05:11 UTC; R25-S01 to R25-S13). TypeSafe's status page lists one new, resolved event: a "System overload" on the us-east-1 API endpoint on 30 Oct, degraded for 1 hour and 15 minutes and resolved at 21:15 UTC (R25-S04). The home page's notice about the jev-latest move ended its seven-day shelf life; the fact stays on versions and aliases. Details: access timeline. Documented
Not checked or not re-read
- Decision-1's input cell (the catalog's text-only sentence), its 32,768-token catalog context and its catalog licence text: the Foundry catalog page builds its text in the browser, so these keep their 13 Oct 2026 dates.
- The graded-test cells (Decision-1's B41 row and the card's graded tests): not re-read; no benchmark was run.
- The contents of Kev
kev-1.0and Layav0.4.18; the Alternatives rows still describe the earlier versions. - Eevo's errors and authentication pages (19 Oct 2026 dates kept); the terms of OpenRouter's Clef hosts (Makora, DeepInfra, PrimeIntellect).
- No request was sent to any provider, and no account was created: every fact here is what the providers' pages say.
Sources
Sources read on 31 Oct 2026 (UTC times)
- R25-S01 to R25-S13: TypeSafe console sign-in pages, the CEO post (X embed endpoint), docs quickstart, status page, models page, images page, homepage, blog index, Jev 1.14 post, Enterprise page, docs changelog address and the OpenRouter guide, 05:11.
- R25-S19 to R25-S30, R25-S49: TypeSafe's legal pages, errors page and question-type pages, and the route listings, 05:11–05:27.
- R25-S38 to R25-S40, R25-S227 to R25-S237: Pydantic AI files at v2.60.0 to v2.64.0 and the v2.64.0 tag page;
@effect/ai-typesafe4.0.1 and 4.1.1 files; TanStack AI at commit 5651e07 and its npm entries, 05:13–05:15. - R25-S42, R25-S43: OpenRouter endpoints for the Decisions API, Decision-1, Clef and Clef-flash; the Vercel AI Gateway model list, 05:14–05:19.
- R25-S200 to R25-S205, R25-S243 to R25-S246, R25-S46: Microsoft Product Terms (MCA and EA/EAS), Foundry model-specific terms, data-privacy and abuse-monitoring pages, the Decision-1 how-to, the Foundry catalog page and the Foundry quotas page, 05:11–05:22.
- R25-S206 to R25-S210, R25-S242, R25-S247 to R25-S249: OpenAI Decisions guide (markdown and HTML), API reference (markdown and HTML), changelog, data-controls page and rate-limits guide, 05:11–05:20.
- R25-S211 to R25-S222, R25-S47: Liquid pricing, d1 docs and launch post; Nace.AI Drex pricing; Mouse pricing, versions, terms, privacy, limits and pricing reference; Eevo pricing and the Eevo announcement; SubQ docs and launch post, 05:11–05:27.
- R25-S238 to R25-S241, R25-S250 to R25-S255, R25-S48: Cloudflare Clef model pages, Workers AI data-usage page, limits and errors pages, and the Clef model cards on Hugging Face, 05:14–05:27.
- R25-S256 to R25-S263: Kev and Laya repository, model-card and latest-release pages, 05:20.
Earlier dated reports: 24–30 Oct 2026; all reports are listed on Research.