Shaduf.Research preview
Jev: Use Cases, Alternatives & Products/Jev limitations and fit check
Start hereJev 1.13 and Jev 1.14 · limits checked · input types and context limits rechecked · rate limits rechecked · image requests in 8 integrations read · models page input, question types and language section read · three image-table rows rechecked

What Jev (TypeSafe AI) cannot do: documented limits and a task fit check

Jev answers typed questions from a fixed answer set. Tasks that need free text, raw images (an opt-in experiment since 14 Oct 2026, not production-ready per TypeSafe), exact arithmetic or an answer outside the list you gave it are a poor fit. This page lists the 17 limits TypeSafe documents, with its workaround for each, has a three-question check you can run on your task, and shows which SDKs and integrations send images to Jev.

Short answer

TypeSafe documents that Jev 1.13 is weak at literal-versus-intended reading, arithmetic and counting, date comparison, multi-step questions, long irrelevant state and adversarial text. It cannot generate text and takes no audio or video. Until at least 2 Nov 2026 its models page called English the primary training language; on 7 Nov 2026 that section is gone and TypeSafe's changelog lists "Better multilingual support" for Jev 1.14, with no multilingual accuracy figures published. Since at least 7 Nov 2026 (first observed 7 Nov, 15:29 UTC) the models page lists the input of both Jev 1.14 and Jev 1.13 as "Text, images", while the Images page still requires an opt-in header and says images are not yet ready for production; Box questions need Jev 1.14. The pool sent no image (earlier reading: images only as an opt-in experiment since 14 Oct 2026, checked 16 Oct 2026, 05:21 UTC). The limits listed here were written for Jev 1.13; TypeSafe's models page gives Jev 1.14, now jev-latest (24 Oct 2026), the same context limits and price, and TypeSafe has not published a separate limits page for 1.14. Keep computation, permissions and actions in your code, and use Jev to pick among options you define, score against a rubric, or estimate a yes probability.

Change, 24 Oct 2026: this answer said "TypeSafe has not published a separate limits page for the Jev 1.14 preview". Since the move first seen on 24 Oct 2026, 05:12 UTC, jev-latest gives Jev 1.14 on TypeSafe's API; the limits table now says its rows 1–10 are TypeSafe's 1.13 findings (R23-S05). The page label drops "preview", and the pydantic-ai-slim row notes the newer 2.61.0 release (not re-read).

Is Jev multimodal?

Text by default; images only as an experiment. Since at least 7 Nov 2026 (first observed 7 Nov, 15:29 UTC; the 2 Nov 2026, 11:13 UTC read still showed "Text") 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". Box questions need Jev 1.14. Images are accepted only when you send the opt-in header TypeSafe-Danger-Zone-Images: true; TypeSafe says images and Box questions "aren't yet up to our standards to call ready for production". So image input is experimental and not production-ready per TypeSafe. There is no audio or video input ("It does not natively accept audio or video (yet)"). (Models page and images page, read 7 Nov 2026, 15:29 UTC, R27-S05, R27-S06; first checked 16 Oct 2026, 05:21–05:22 UTC.) Documented

TypeSafe's post on Jev 1.14 says jev-1.13.0 "was multimodal all along". Reported No image was sent here, so image quality is not tested.

pydantic-ai-slim 2.56.0 and 2.57.0 reject a request with images before any call is made (UserError('TypeSafe does not support image input.'), models/typesafe.py L219–220; package files read 16 Oct 2026, nothing run). So the experimental image input is not reachable through that package. For comparison, OpenAI's Decisions API and Cloudflare's Clef take images; Microsoft-Decision-1 is text only (OpenAI and Microsoft read 13 Oct 2026, Cloudflare 12 Oct 2026; see the input row of the decision-API comparison and their rows on Alternatives). More on the 1.14 changes: what changed with Jev 1.14.

Which SDKs and integrations send images, and who adds the header: Jev with images (8 integrations read in package files, 23 Oct 2026; pydantic-ai-slim still refuses images at 2.60.0). On 23 Oct 2026 the models page still lists the input as "Text" while the images page says "Jev 1.13 and Jev 1.14 both accept images"; both are quoted there. On 7 Nov 2026 the models page also lists "Text, images" for both versions, so the two pages now agree that images are accepted.

Change, 16 Oct 2026: this box answered "No" until 14 Oct 2026, 16:46 UTC, when the models page said "Text only … No image, audio, or video input" and advised turning images, audio or video into text or structured fields first. On 16 Oct the page says "Text" and documents images as an opt-in experiment.

Change, 7 Nov 2026: the models page lists the input of both versions as "Text, images" (first observed 7 Nov 2026, 15:29 UTC; "Text" on 2 Nov 2026). The images page still calls images and Box questions not yet ready for production and still requires the opt-in header (R27-S05, R27-S06). No image was sent here.

Sources: TypeSafe models page (R17-S05; earlier R15-S06), images page (R17-S06) and Jev 1.14 post (R17-S09); pydantic-ai-slim 2.56.0 and 2.57.0 models/typesafe.py (R17-S58; earlier read at 2.55.0, R13-S119); OpenAI Decisions guide (R14-S219) and API reference (R15-S214); Microsoft Foundry catalog (R15-S44); Cloudflare Clef docs (R14-S400 to R14-S402). Models named "Jev … VL" on Hugging Face are third-party models, not TypeSafe's: see named like Jev.

Jev with images: does your SDK or integration send them?

Key finding

Of 8 integrations read, 5 send images to Jev and 2 of those add TypeSafe's opt-in header for you; pydantic-ai refuses images, and both LangChain packages pass them as text.

Documentedpackage files and TypeSafe's docs, read ; nothing was installed or run and no image was sent

Images are still an opt-in experiment that TypeSafe says is not ready for production. "Send images to Jev" means the integration puts the image into the request; it does not mean the image was tested or that answers on images are reliable.

What 8 integrations do with an image you give them (read in package files, 23 Oct 2026)

  • Sends the image, adds the header for you (2): @ai-sdk/typesafe-ai 3.0.24, TypeSafe's n8n node 0.10.0
  • Sends the image; you set the header (3): @typesafe-ai/sdk 0.7.0, typesafe-sdk 0.7.5, LiteLLM 1.106.1
  • Refuses images before the call (1): pydantic-ai-slim 2.60.0
  • No image type; the image goes to Jev as JSON text (2): @langchain/typesafe 0.0.2, langchain-typesafe 0.0.2 (read at 0.0.1a3; unchanged at 0.0.2)

One block per integration read (5 + 1 + 2 = 8). Shades only separate the four groups; they are not the failure-map colours.

8 integrations read: 5 pass images, 1 refuses them before the call and 2 have no image type. Of the 5, 2 add the opt-in header for you and 3 need you to set it.
Integration (version read)Passes imagesOpt-in headerWhen the image path failsWhere read
TypeSafe API (docs, 23 Oct 2026)Yes: PNG or JPEG, inline base64 or multipart; no URL field documentedYou set it: an HTTP header on each request with an image or a Box questionAPI error to the caller: 400 without the header or for a rejected image, 413 over a size limit; the whole request failsimages.md L19, L327–335, L371–378 (R22-S100)
@ai-sdk/typesafe-ai 3.0.24 (same source at 3.0.26) @ai-sdk/typesafe-ai 3.0.30 was released on 29 Oct 2026 (17:10 UTC; patch releases 3.0.27 to 3.0.30 since 24 Oct); this site last read the provider at 3.0.26; not re-read (R24-S35). 3.0.31 followed on 30 Oct 2026 (15:39 UTC; patch); newer version exists, not re-read (R25-S36). 3.0.34 (5 Nov 2026) is newer than the version this page describes; not re-read (R27-S25). 3.0.35 (7 Nov 2026), compared with 3.0.34 on 9 Nov 2026: it now rejects empty Choice or Score criteria and empty Score levels before sending, sends primitive JSON state as text, and returns Score legends in provider metadata; its image lines are not in that diff (R29-S18). The versions between 3.0.26 and 3.0.34 were not compared line by line.Yes: PNG or JPEG file parts, inline dataSets it for you (any image part or Box question)Rejected before the call (other media types, URL data); API error to the caller for size or dimensionssrc/typesafe-ai-decision-model.ts L148–158, L216, L245–248 (R22-S106, R22-S116)
@typesafe-ai/sdk 0.7.0Yes: any bytes; a file that is not PNG or JPEG is sent as generic binaryYou set it (defaultHeaders or request headers)Rejected before the call (a URL string, empty data); otherwise API error to the callerdist/index.mjs L226–264, L258, L275 (R22-S108)
typesafe-sdk 0.7.5 (Python)Yes: bytes, a path, base64 or a Pillow image; a file that is not PNG or JPEG is sent as generic binaryYou set it (extra_headers)Rejected before the call (a URL string, empty data, invalid base64); otherwise API error to the caller_core/assets.py L24–28, L164, L262, L269 (R22-S109)
pydantic-ai-slim 2.60.0 (TypeSafe files as in 2.59.0) pydantic-ai-slim 2.61.0 was published on 24 Oct 2026 (03:55 UTC); not re-read (R23-S41). pydantic-ai-slim 2.63.0 was released on 28 Oct 2026 (02:03 UTC; 2.62.0 on 27 Oct); newer version exists, not re-read; the TypeSafe model file was last read at 2.60.0 (R24-S36). pydantic-ai-slim 2.64.0 (31 Oct 2026, 00:22 UTC): the TypeSafe files were compared with 2.60.0 for the failure path on 31 Oct 2026; the changes found are an import (L8) and a new error branch (L252–254); the image cells stay as read at 2.60.0 (R25-S36, R25-S38). pydantic-ai-slim 2.66.0 (4 Nov 2026): TypeSafe files identical to 2.64.0 by file hash, 8 Nov 2026 (R28-S215).No: refuses a request with images before the call (UserError)You can set it (extra_headers), but images are refused either wayRejected before the callmodels/typesafe.py L222–223 (R22-S112, R22-S118)
TypeSafe's n8n node 0.10.0 (ec81714)Yes: PNG or JPEG from the item's binary dataSets it for you whenever an image is attachedRejected before the call (missing binary, wrong type, bad reference); API error to the caller for size (the node does not check size)api.ts L125–131, L141–148; TypeSafeAi.node.ts L76–131 (R22-S206, R22-S208)
LiteLLM 1.106.1 (TypeSafe files the same at 1.106.2), decisions routeYes, only with the header: PNG or JPEG; image URLs are fetched by LiteLLM and sent inlineYou set it (extra_headers); LiteLLM checks for it and forwards itRejected before the call (no header, another type, an unreachable URL); with drop_params on, the image is dropped and the request is still sent, so Jev decides without itllms/typesafe/decisions/transformation.py L21–40, L81–106; decisions/call.py L377–387, L486–495 (R22-S115, R22-S119)
@langchain/typesafe 0.0.2No image typeCannot be setThe request is still sent: the image goes to Jev as JSON text, so Jev decides without seeing itdist/utils/messages.js L47–58 (R22-S113)
langchain-typesafe 0.0.2 Image path read at 0.0.1a3 (23 Oct 2026). Re-read at 0.0.2 (wheel, 8 Nov 2026): failure path unchanged; adds an optional classifier= argument on both experimental middlewares; _state.py is byte-identical, so the image handling is unchanged (R28-S212, R28-S213).No image typeCannot be setThe request is still sent: the image part goes to Jev as JSON text, so Jev decides without seeing itlangchain_typesafe/_state.py L15–52 (R22-S114)

This table is not part of the failure map: it records what each integration does with an image, read in its files. No image was sent. Where an image is dropped without an error (both LangChain packages always; LiteLLM only with drop_params on), your code gets an answer that was made without the image. Build pages: TypeScript, Python, frameworks, n8n.

What TypeSafe's docs say about image requests

Images page and models page read 23 Oct 2026, 05:19 UTC (re-read 05:26 UTC, byte-identical); API reference read 05:23 UTC. Documented

Status
"Images and Box questions aren't yet up to our standards to call ready for production, but we're excited for dreamers and builders to play with them. … The request format, the answers, and the limits on this page can change between releases" (images page).
Opt-in header
"Images and Box questions are off by default. To use them, send the TypeSafe-Danger-Zone-Images: true header with each request that contains an image or a Box question." The API reference lists it as an HTTP header parameter.
Which versions
The images page says "Jev 1.13 and Jev 1.14 both accept images. Box questions need Jev 1.14." On 7 Nov 2026 the models page lists the input of both versions as "Text, images" ("It also reads PNG and JPEG images anywhere in the request. It does not natively accept audio or video (yet)") and its "Question types" row lists Box for Jev 1.14 only (R27-S05). Until 2 Nov 2026 the models page listed the input as "Text", which did not match the images page; first observed matching 7 Nov 2026.
Formats
"png" | "jpg" | "jpeg"; "PNG or JPEG, still images only. Animated PNGs are rejected." The format is detected from the bytes.
How an image is sent
Inline data_base64 ("no data: prefix") or as a multipart file part. An image URL field is not stated on the images page, the errors page or the API reference (05:23–05:26 UTC); the asset schema has type, format, data_base64 and part only.
Limits
16 MiB per image file; 280 to 1568 pixels per side; 78,400 to 2,458,624 pixels in total; request body 32 MiB; all images 32 MiB after processing; expanded request 32 MiB. "An image outside them is rejected, never resized to fit. If any image in a request is rejected, the whole request fails: there are no partial answers." No fixed image count is given.
Errors
400 without the header ("Images are not turned on"; no error name given); 400 for a rejected image or asset (not PNG or JPEG, damaged, animated, too small, invalid base64 or format, an unreferenced asset); 400 max_tokens_exceeded when images and text exceed the context length; 413 over a size limit. The errors page does not mention images.
Billing
"Each image counts as input tokens, billed at the same per-token input price as text. There is no separate charge for images."

Sources: TypeSafe images page L9, L19, L61, L79–83, L158, L183, L327–337, L365, L371–378 (R22-S100); models page L16, L22 (R22-S101); API reference, System One L46–53, L371–372 (R22-S103); request rules and errors (R22-S102).

A route that lists image input is a listing; it does not show that the route forwards TypeSafe's opt-in header. Each route's image field, read 23 Oct 2026, is on versions and aliases; whether OpenRouter, Vercel, Opper, Netlify or Cloudflare forward or add the header is not known.

Other projects that send images to Jev (leads, by date)

  • 22 Oct 2026: jev-gateway (vinilana/jev-gateway) merged PR #67 (83b5b98); it is not in an npm release (latest 0.5.1). It sets the opt-in header itself; images that are URLs, not PNG or JPEG, or outside the size limits are replaced by a text placeholder and the request is still sent (src/images.ts L55–65, L142–195; R22-S233, R22-S238). Its row on the failure map was read at npm 0.5.1 (37ff982), before this change.
  • 20 Oct 2026: Jev Cards (Jose-Sabater/jevcards, 8e67173) and Photo Ferret (JonesSteven/photo-ferret, 7949f80), both shown on Hacker News, send images to Jev and set the header themselves (server/jev.ts L137; server/jev.ts L39–44). Code read; live operation not verified (R22-S261, R22-S269, R22-S262, R22-S266).
  • 19 Oct 2026: JevLight (usail-hkust/JevLight) merged PR #2 (ff0395f), a research integration that sends camera images as multipart, sets the header itself, and rejects images before the call on a transport other than TypeSafe's (jevlight/jev_client.py L216–217, L530–538; R22-S250, R22-S254).
  • 18 Oct 2026: racecraft-lab/racecraft-plugins-public #1377, an open bug report: an image reference token can collide with the request text (Go file not read; Reported; R22-S245).
  • 17 Oct 2026: HRyohni/jev-front-back-id-matching, an evaluation script that sets the header itself (0633d2e); its README results are the author's own (Reported; R22-S258, R22-S260).
  • 15 Oct 2026: microsoft/agent-framework #9410, an open feature request saying the TypeSafe connector "rejects every non-text message content" and "never sends" the header (connector source not read; Reported; R22-S244).
  • Not located: a "hibana #83" image lead (hibanaworks/hibana issue, pull request and discussion 83 returned HTTP 404 on 23 Oct 2026; R22-S246 to R22-S248). Seen 21 Oct 2026, not read: anorthernman-developer/jev-thumbnail-critic, Luminary-AI-Solutions/image-moderation-bench, harigittihub/SentinelEdge-a-two-speed-AI-inspector.

Not verified about images

  • No image was sent and no Jev call was made; every row above is a reading of TypeSafe's docs or of package files at the named version.
  • Whether any route (OpenRouter, Vercel, Opper, Netlify, Cloudflare) forwards or adds the opt-in header.
  • How LiteLLM reports a 413, and its chat path; the exact JSON form of an image in langchain-typesafe; whether a malformed assets object returns 400 or 422.
  • The n8n node's missing-binary error was read in n8n 2.44.3; your n8n version may differ.
  • How accurate Jev is on images: no graded study on the benchmarks ledger tests image input, and TypeSafe says accuracy "will change as we update the models".

Change, 9 Nov 2026: @ai-sdk/typesafe-ai 3.0.35 compared with 3.0.34 (image lines not in the diff; it adds local checks for empty criteria and Score levels) (R29-S18). No integration cell changed.

Change, 8 Nov 2026: langchain-typesafe re-read at 0.0.2 (image handling unchanged; the row now names 0.0.2); pydantic-ai-slim 2.66.0 TypeSafe files identical to 2.64.0; @ai-sdk/typesafe-ai 3.0.35 noted, not re-read (R28-S47, R28-S212, R28-S213, R28-S215). No integration cell changed.

Change, 7 Nov 2026: newer releases noted in three rows: @ai-sdk/typesafe-ai 3.0.34 (not re-read), pydantic-ai-slim 2.66.0 and langchain-typesafe 0.0.2 (both read on 8 Nov 2026) (registries, 7 Nov 2026; R27-S25). TypeSafe's models page now lists the input of both versions as "Text, images" (first observed 7 Nov 2026; R27-S05); the images page still requires the opt-in header. No integration cell changed.

Change, 31 Oct 2026: newer releases noted in two rows: @ai-sdk/typesafe-ai 3.0.31 (30 Oct 2026; patch; not re-read) and pydantic-ai-slim 2.64.0 (31 Oct 2026; TypeSafe files compared for the failure path only) (registries, 31 Oct 2026, 05:14 UTC; R25-S36, R25-S38). No reading in this table changed.

Change, 30 Oct 2026: newer releases noted in two rows, not re-read: @ai-sdk/typesafe-ai 3.0.30 (29 Oct 2026; patch releases) and pydantic-ai-slim 2.63.0 (28 Oct 2026; minor) (registries, 30 Oct 2026, 05:20 UTC; R24-S35, R24-S36). No reading in this table changed.

Change, 23 Oct 2026: section added (TypeSafe's image docs; 8 integrations read in package files; leads by date).

Fit check: does your task fit?

What the three questions test

QuestionWhy it mattersTypeSafe's own guidance
1. Can you list the possible answers in advance?Choice takes up to 255 named options, Score an ordered rubric of at least 2 and at most 10 levels, Noul a yes/no question. (Box, on Jev 1.14 only and behind the images opt-in header, returns a bounding box in an image; models page, 7 Nov 2026.) Jev cannot return an answer you did not offer. Arithmetic and date comparison belong in code, not in the question."System One models do not write replies, produce code, or generate explanations of their reasoning. You define the possible answers through primitives." System One concepts
2. Will your own code decide what happens after the answer?A valid answer can still be wrong. In the audited projects the action policy lives in the calling code and differs a lot between projects."code owns the workflow and AI handles narrow, structured decisions … Keep control flow, deterministic rules, and side effects in code." How to build with System One
3. What happens when Jev is unsure, wrong or unavailable?Errors happen (401, 422, 429, 529 are documented). One audited trading gate lets an entry through when Jev fails; a moderation plugin holds the comment instead. Of six moderation projects read on 2 Oct, four leave the post up when Jev fails (content moderation). For a Noul, the middle band of probabilities is the "unsure" zone."Low confidence: Do not act. Route to a human, request clarification, or fall back to a different system." Confidence page

Documented Quotes checked 28 Sep 2026. The check runs entirely in your browser: it sends nothing, calls no model and stores nothing. It gives a starting recommendation, not an accuracy estimate. To measure accuracy on your own cases, see Confidence thresholds; for real examples, see the failure matrix.

Documented limits and workarounds (Jev 1.13)

Rows 1–10 come from TypeSafe's model jaggedness page ("Applies to jev-1.13. Last reviewed 2026-09-17"; section numbers shown). Rows 11–17 come from the Models page, the API reference and the primitives page. All were checked on 28 Sep 2026. Documented These are failure modes TypeSafe documents, not error rates anyone measured.

Requests to jev-latest on TypeSafe's API now get 1.14; TypeSafe's models page links only a Jev 1.13 jaggedness page (24 Oct 2026, 05:12 UTC; R23-S05), so rows 1–10 are TypeSafe's findings for 1.13. Whether they also hold for 1.14 is not published and not tested here.

#LimitTypeSafe saysConsequence for youDocumented workaroundSource
1Literal reading"answers the question you wrote, not the one you meant"Scoping words and negations are taken at face valueState the exact condition; put boundary cases in the criteria; split an interpretive question into two literal ones and combine them in code§1
2Arithmetic and counting"Jev is not a calculator"; "does not count reliably"Sums, counts and numeric comparisons are unreliableDo arithmetic in code; ask one question per item and add up in code; turn numbers into named ranges§2
3A Score is not a number"score levels are weak in numerical calibration"You cannot rebuild a magnitude from scoreCompare score with a threshold; do not interpolate between levels§2
4Dates and times"reads dates as text, not as ordered quantities"Ordering, durations and date windows are unreliableExtract date parts with a Choice (with a "not stated" option) and compare in code; see the date-extraction cookbook§3
5Indirect questions"multiple hops of reasoning costs accuracy"Double negatives and property-of-a-property questions get worse answersAsk directly and name the relevant field of the state§4
6Large, irrelevant state"Accuracy falls as the state grows with content unrelated to the decision"Long context lowers accuracy and makes errors harder to traceRetrieve and filter in code first, or filter passages with a Noul; see the RAG passage cookbook§5
7Adversarial content"does not treat it as hostile by default"Injected or self-arguing text in the state can move the answerWrite explicit criteria and test edge cases before deployment§6
8Contradictory instructions"might get confused"Criteria that invert the instruction (for example in a Noul) perform worseKeep criteria consistent with the instruction§7
9No structural guaranteesA question and its negation need not agree (0.72 + 0.47 = 1.19)Arithmetic across questions and moving a threshold between question types are invalidAsk each decision one way; enforce identities in code§8
10No text generation"not trained to generate text"No free-form extraction, summaries, replies or codeProduce candidates with a regex or an LLM, then let Jev pick one with a Choice§9
11Text input by default; images experimentalInput "Text, images" (both versions; models page, 7 Nov 2026); images only with the opt-in TypeSafe-Danger-Zone-Images header, "aren't yet up to our standards to call ready for production" (images page, 7 Nov 2026); Box questions need Jev 1.14; no audio or video
Change, 7 Nov 2026: the models page input row reads "Text, images" (first observed 7 Nov; "Text" from 16 Oct to 2 Nov 2026).
Change, 16 Oct 2026: until 14 Oct the page said "No image, audio, or video input."
Audio and video need a separate conversion step; images can be sent only as an experiment that TypeSafe says is not production-ready; which SDKs send them: Jev with imagesConvert to text or structured fields first, and check that step separatelyModels page; images page
12Context limits64k tokens per request; 32k for state plus the longest question (default, both versions). The opt-in TypeSafe-Danger-Zone-Long-Context header raises both to 262k at the same price per token; TypeSafe says it is "slower and less accurate" and not yet production-ready (16 Oct 2026)
Change, 16 Oct 2026: the 262k opt-in is new; the 64k / 32k default is unchanged since 28 Sep.
Long documents must be cut or filteredFilter the state; ask several questions about one state in one requestModels page
13LanguagesUntil 2 Nov 2026 the models page said "English is the primary training language… Other languages, including CJK scripts, are handled but not equally well". On 7 Nov 2026 that section is gone; TypeSafe's new model changelog lists "Better multilingual support" for Jev 1.14 (dated October 14, 2026) (R27-S05, R27-S11)No multilingual accuracy figures were found on the models page or changelog (7 Nov 2026)"test on your own content"; watch confidence when routing non-English text (language handling in moderation projects: content moderation)Models page; model changelog
14Closed answer setA Choice picks only from the options you give (at most 255)It cannot propose a new class, and it always picks oneAdd an "other" or "none of the above" optionAPI reference
15Questions are independent"one answer does not become context for another question"No chaining within one requestMake a second request from your code when one answer depends on anotherPrimitives page
16Rate limits are not fixed or published"Rate limits are set per organization"; current limits are in the developer console (16 Oct 2026; unchanged on 30 Oct 2026, 05:14 UTC, R24-S05). Until 14 Oct: limits "can change without notice"
Change, 30 Oct 2026: TypeSafe now takes requests for higher limits through a rate limit increase request form (models page, checked 30 Oct 2026, 05:14 UTC; R24-S05); the change happened between 24 Oct 2026, 05:12 UTC and 30 Oct 2026, 05:14 UTC. Limits are still set per organization and shown in the developer console.
Capacity planning is uncertain, and the numbers that apply to you are visible only in your console. The published numbers changed twice: between 29 Sep, 19:24 UTC and 30 Sep, 20:47 UTC from 250,000 tokens/s and 1,200 requests/min to 100K tokens/s and 40 requests/s, and between 2 Oct, 05:14 UTC and 3 Oct, 05:16 UTC to 100K tokens/s and 80 requests/s (last seen 14 Oct 2026, 16:46 UTC). By 16 Oct 2026, 05:21 UTC the models page listed no numbers, and still listed none on 24 Oct 2026, 05:12 UTC or on 30 Oct 2026, 05:14 UTC (errors and rate limits)Back off and retry; enterprise plans for guaranteed limits; request a higher limit through TypeSafe's rate limit increase request form (models page, checked 30 Oct 2026, 05:14 UTC); see errors and rate limitsModels page
17No per-customer tuning"not fine-tuned or LoRA-adapted with customer data"Domain adaptation happens only through state, instructions and criteriaPut rules in the criteria; split decisions; train your own model on Jev outputs (AutoResearch cookbook)Models page

Reported, not documented: text in state that contains a shell curl … https:// command has been blocked by a Cloudflare HTML 403 before reaching the API (reported 23 Sep, still open). This matters for coding-agent guardrails; see errors and rate limits. Reported

When to use something else

  • You need generated text or code: an LLM, optionally with structured output.
  • You have thousands of labelled examples and a stable label set: a trained classifier may be cheaper and local; see alternatives.
  • You need the model to run on your own hardware: independent open-weight models, with their own licences and accuracy; see alternatives.
  • The rule can be written exactly (a regex, a threshold, a lookup): write the rule.

For a full comparison with 10 criteria (labelled data, changing labels, text output, latency, residency, probabilities, input length, arithmetic and language, volume, vendor risk) and the published head-to-head results, see Jev vs an LLM vs a classifier: when to use which. What published tests found for each task type is on the benchmarks ledger.

What was not verified

  • This pool did not test any limit by calling Jev; every row is TypeSafe's own documentation.
  • How often each weakness occurs on a real task is not measured here. The only published non-English test with a baseline is Bryo's German and English email benchmark (grade C, B12). TypeSafe's jaggedness page still said "Last reviewed 2026-09-17" on 30 Sep 2026.
  • Gateway-specific limits beyond the listed 32k context were not checked.

Search published pools, pages, reports, and evidence.