Jev 1.14.0 danger zone flags: images, Box and long context, and which belong in production

TypeSafe AI shipped Jev 1.14.0 on 14 October 2026 with three opt-in "danger zone" features — images, a Box type and long context — that the company says it cannot vouch for in production, and an independent board scored 1.14 with images 10 points below 1.13 with images on the same 600 items. Because TypeSafe's client SDKs default to the `jev-latest` alias, which moves from 1.13.0 to 1.14.0 around 21 October, the release post's guidance is to pin `jev-1.13.0` and keep all three headers out of production, with long context the sole exception for requests that overflow 64k tokens once accuracy has been measured. The release also added an organization-wide console setting, "Model improvement," that lets TypeSafe "read and analyze" inputs and outputs while stating it "will still not train on your queries.

Jev 1.14.0 danger zone flags: images, Box and long context, and which belong in production As of 19 October 2026, keep Jev https://stackness.dev/tools/typesafe-jev 1.14.0's three danger zone flags out of production and pin jev-1.13.0 before jev-latest moves to 1.14 around 21 October. TypeSafe AI shipped the release on 14 October with images, a Box type and long context behind headers it says it cannot "vouch for" in production, and an independent board scored 1.14 with images 10 points below 1.13 with images on the same 600 items. The one exception is long context for the few requests that overflow 64k tokens, once you have measured accuracy on them. A preview model and three opt-in headers are at stake Jev 1.14.0 is a preview model plus three features that each need a request header, and the decision is which of them a production system should send. Nothing is on by default. The risk that needs action this week is the alias: TypeSafe's client SDKs default to jev-latest , which moves from 1.13.0 to 1.14.0 a week after release. The release post https://typesafe.ai/blog/jev-1-14 calls 1.14.0 "small, subtle improvements to everything" with "the largest gain in multi-lingual support". TypeSafe's model docs https://docs.typesafe.ai/models aliases spell out what an alias move means: answers "can change without any change to your code", and teams that tuned confidence thresholds should send the versioned ID. Thresholds tuned on 1.13 have not been checked on 1.14 until you check them. The release also added an organization-wide console setting, "Model improvement", that lets TypeSafe "read and analyze" your inputs and outputs. TypeSafe says it "will still not train on your queries". It is a data decision, not a model one, and it applies to the whole organization at once. What is a danger zone flag in Jev? A danger zone flag is a request header that turns on a Jev feature TypeSafe ships but will not vouch for in production. The release post says these features "will be here to stay for future models" but "we think there is still way too much jaggedness for us to vouch for these for production use cases". Without the header, an image or Box request returns 400. The docs' callout on the images and Box pages goes further: "The request format, the answers, and the limits on this page can change between releases, and accuracy will change as we update the models." No notice period is stated for those changes. The only deprecation promise is for whole models, and it is an intent: 60 days' warning before 1.13.0 is deprecated and 30 for 1.14.0, under a policy TypeSafe says it will publish. On 19 October that policy page returned 404. | Feature | Header | Models | What TypeSafe says | Price | Limits | |---|---|---|---|---|---| | Jev 1.14.0 | none; model: "jev-1.14.0" or jev-preview | 1.14.0 | Preview; becomes jev-latest about 21 October | $0.042 per million input tokens, output free | 64k per request, 32k per question | | Images | TypeSafe-Danger-Zone-Images: true | 1.13.0 and 1.14.0 | "Dreamer-ready"; models "perform worse on multimodality" | Image tokens at the input price; no estimate published | PNG or JPEG, 280 to 1,568 px a side, 16 MiB a file | | Box | the images header | 1.14.0 or jev-preview only | "There is no recommended threshold yet" | As images | One box per question, edges from 0 to 1, no "not found" answer | | Long context | TypeSafe-Danger-Zone-Long-Context: true | 1.13.0 and 1.14.0 | "slower and less accurate" | Same per token | 262k per request and per question | Sources: images https://docs.typesafe.ai/images , Box https://docs.typesafe.ai/primitives/box , models https://docs.typesafe.ai/models , read on 19 October 2026. Benchmark Heaven noted that the docs gave a 256 px minimum side when it tested, while the API rejected anything under 280, which is what the docs now say. Jev 1.14 scored lower than 1.13 on an independent image board Benchmark Heaven ran ImageJevBench https://benchmarkheaven.com/image-jev-bench on both versions with images on 14 October: 600 items, the same harness, one choice question each, no retries. Jev 1.13.0 scored 59.9 on intelligence and 85.0 on calibration. Jev 1.14.0 scored 49.9 and 70.5. Both cost about $0.16 per 1,000 image decisions, with median answers in 0.378 and 0.393 seconds. That is one third-party pass on one item pool, and it runs against TypeSafe's own chart, which puts 1.14.0 ahead of 1.13.0 overall, 60.0 against 58.9, under a banner that reads "Please don't trust them". On text, the same publisher's JevBench https://stackness.dev/tools/jevbench board has 1.14.0 at 59.9 against 63.6 for 1.13.0, on different item draws it describes as equated. The benchmarking post https://stackness.dev/blog/how-do-you-benchmark-a-system-one-model-what-jevbench-scores-and-what-calibration-says-it-cannot covers what those scores measure. The practical reading: if you prototype with images, prototype on jev-1.13.0 with the images header, and do not assume 1.14 is the better vision model. Box always returns a box, even when the object is absent The Box type returns one bounding box, {"top":0.25,"bottom":0.82,"left":0.44,"right":0.84} with a confidence, for an object named in the question. The docs https://docs.typesafe.ai/primitives/box say a Box answer "is always a box. There is no 'not found' answer", and recommend pairing each Box with a yes or no question a Noul that asks whether the object is there at all. On Benchmark Heaven's Box track, Jev 1.14.0 reached a mean intersection-over-union score of 62.8, with 70% of boxes at 0.5 or better, at 0.33 seconds and $0.057 per 1,000 boxes. It ranked seventh of 18 on a score that averages box accuracy with calibration; on box accuracy alone, 13 of the 18 drew better boxes, some of them in under two seconds and H2O Lightning 4B in 0.35. Jev was weakest on small targets, at 47.2. Long context is slower and less accurate, and OpenRouter turns it on for you Long context raises both Jev limits to 262k tokens at the same price per token, and TypeSafe's docs say it is "slower and less accurate than staying within the default limits, so turn it on only for requests that need it". I found no published latency or accuracy measurement for it. A full 262k request costs about $0.011 at list price, against about $0.0028 for a full 64k one. Jev 1.13's own jaggedness page https://docs.typesafe.ai/model-jaggedness/jev-1.13 already warns that "Jev suffers from context rot, so unrelated material in the state costs you accuracy." Long context makes that easier to hit. The flag can also be on without anyone sending it. OpenRouter's Jev guide https://openrouter.ai/docs/guides/community/jev says requests with more than about 32,000 bytes of text, and every request with an image, "run in TypeSafe's long-context mode". A team that reaches Jev through OpenRouter https://stackness.dev/tools/openrouter is in the danger zone on those requests already. OpenAI Decisions takes images in beta; Microsoft-Decision-1 is text only The OpenAI Decisions API https://stackness.dev/tools/openai-decisions-api accepts images as part of its public beta with no extra flag, has no Box type and costs $0.10 per million input tokens guide https://developers.openai.com/api/docs/guides/decisions ; it runs on GPT-6 Luna, whose model page https://developers.openai.com/api/docs/models/gpt-6-luna lists a 1,050,000-token context and doubles the input price above 272k tokens. Microsoft-Decision-1 https://stackness.dev/tools/microsoft-decision-1 "does not accept or produce image, audio, or video content" and has a 32,768-token window model card https://ai.azure.com/catalog/models/Microsoft-Decision-1 . On ImageJevBench, GPT-6 Luna through Decisions scored 45.0 and 80.7, below Jev 1.13 with images on both counts and above Jev 1.14 with images on calibration. Microsoft's model is not on the image board. The earlier comparisons with OpenAI Decisions https://stackness.dev/blog/openai-decisions-api-vs-jev-price-per-decision-latency-and-switching and Microsoft-Decision-1 https://stackness.dev/blog/microsoft-decision-1-vs-jev-price-per-decision-latency-and-switching cover price per decision and latency on text. Among the language models developers pick for their agents https://stackness.dev/categories/llms , these three are the typed-decision APIs in the Stackness catalog. Test a flag on a slice of traffic before it decides anything Danger zone headers are sent per request, so a flag can run on a copy of production traffic while the answer that counts still comes from the pinned model. TypeSafe gives the pieces: pin a versioned ID, log the model field each response returns, and switch when you are ready. Retuning thresholds on the new version is the step that makes the switch safe. There is no official evaluation harness; the console playground and third-party boards are what exists. A test that holds up: 1. Pin model: "jev-1.13.0" in every production call today. 2. Freeze a labelled set from your own traffic, a few hundred items, including the hard ones. 3. Send the same requests to the candidate 1.14.0, or the flag as a shadow call, and log answers, probabilities and confidence side by side. 4. Compare error rate at your current threshold, and how often confidence crosses it, not average score. 5. Switch only the decisions where the candidate wins, with the version and header set per call. Prompt wording moves Jev more than many teams expect. Zach Landes's trolley problem test https://zachlandes.com/trolley-problem/ on 1.13 found that changing only the order of the descriptions changed Jev's choice in 370 of 2,620 pairs. Swift of Major League Hacking cut https://dev.to/theycallmeswift/i-got-jev-to-zero-mistakes-im-still-using-flash-lite-2mo7 false positives from 35 runs to 0 by rewriting one criterion sentence. A published policy and a second image benchmark would change this Three things would change the recommendation. A published deprecation and change policy for danger zone features would let a team plan around them. A second independent image benchmark, or a 1.14 point release, that put 1.14 at or above 1.13 on images would move the image prototype to 1.14. And a published latency and accuracy measurement for long context would show whether it can carry routine traffic. The flags themselves are not the problem. TypeSafe has said plainly what they are, and per-request headers make them easy to contain. What is missing is a promise about how they will change. What I would do - If you run Jev in production on SDK defaults, pin jev-1.13.0 now and test 1.14.0 on a frozen set before 21 October. - If your inputs sometimes exceed 64k tokens, send the long-context header only on those requests, after measuring accuracy on them; trim the state first, since context rot costs accuracy at any length. - If you call Jev through OpenRouter, check which requests cross 32,000 bytes or carry images: those already run in long-context mode. - If you need images or Box, build the prototype on jev-1.13.0 for images, keep Box behind a Noul that asks whether the object exists, and keep both off the decision path until TypeSafe drops the danger zone label. - If your data cannot leave your own control, leave the "Model improvement" setting off; it applies to the whole organization. New posts by email About one email a week with what we published. No account needed. Which of these do you use? Nobody on Stackness runs JevBench yet. Tick what you use and be the first. Play Lonny Eagle Amateur Skater 4A pixel art skateboarding game. Signing up opens it straight away.Sign up to play https://stackness.dev/games/lonny-eagle