Jev AI

Jev version guide · release notes read October 15, 2026

Jev 1.14.0

The current Jev release: what TypeSafe changed, the three experimental features that shipped with it, the jev-latest switch, and how to call it.

TypeSafe released jev-1.14.0 on October 14, 2026 under the jev-preview alias, and jev-latest now points to it too. For text decisions it is the same hosted System One model with small improvements, the largest in languages other than English. Alongside it TypeSafe opened three "danger zone" features behind request headers: images anywhere in a request, a box question type for locating something in an image, and a 262k-token context. This page collects every published fact about the version in one place, with the date it was read.

jev-latestresolves to 1.14.0, as does jev-preview; pin jev-1.13.0 to keep the previous release
60.0intelligence on TypeSafe’s internal measure, up from 58.9 for 1.13.0
262ktokens with the long-context header, up from 64k
$0.042per million input tokens, unchanged; output free

Try it on this page

Run Jev 1.14 right here

Edit the example or paste your own text, then run it. Sign in to run: new accounts get 20 welcome credits, and a failed run costs nothing.

Open the full playground
Run with

Photo of an invoice · Jev 1.14

1 Input

A supplier sent this scan instead of a PDF: `scan`

2 Questions

  • Yes / No Every line of the total is readable without guessing.
  • Choice What kind of document is this?
  • Score How badly does this need to be re-sent as a PDF?
  • Box the total amount in `scan`

Version, aliases and lifecycle

From TypeSafe’s release post of October 14, 2026 and its model documentation.

Version
jev-1.14.0, released October 14, 2026
Aliases
jev-latest and jev-preview both resolve to jev-1.14.0, on TypeSafe’s models page and on Jev AI; jev-1.13.0 stays available pinned
What changed
TypeSafe describes the release as small, subtle improvements to everything, with the largest expected gain in multilingual support. No flagship capability change for text decisions
Developer
TypeSafe
Model class
System One: typed decisions, no text generation
Lifecycle
jev-1.13.0 stays available; TypeSafe intends 60 days of warning before deprecating it and 30 days for 1.14.0, with a written deprecation policy to come

The danger zone: images, box and long context

Opt-in request headers on the TypeSafe API. TypeSafe’s own words: they are better than what exists, will keep improving, and are still too jagged to vouch for in production.

Images
Header TypeSafe-Danger-Zone-Images: true. Images go in a top-level assets map (PNG or JPEG, base64 or multipart) and are referenced by key from the state, the instructions or the criteria. Works on jev-1.13.0 as well as 1.14.0; billed at the text input-token price
Image limits
16 MiB per image, 256 to 1,568 pixels per side, 78,400 to 2,458,624 total pixels, still images only; 32 MiB per request before and after the references are expanded
Box type
A fourth question type, "box", returns the smallest rectangle around a target in an image as top, bottom, left and right fractions of the image plus a confidence. Only jev-1.14.0 answers it, and it needs an image in the same request. On Jev AI the playground offers it once jev-1.14.0 is selected and draws the rectangle on your picture
Long context
Header TypeSafe-Danger-Zone-Long-Context: true raises both limits from 64k per request and 32k per question to 262k tokens. Same price per token; TypeSafe says it is slower and less accurate than staying within the defaults. Send "long_context": true in a Jev AI API request to opt in; the playground caps the state well below the default limit, so it has no switch
Status
TypeSafe calls these "dreamer-ready", not production-ready, and does not vouch for them in automation. Jev AI forwards all three for the pinned versions jev-1.14.0 and jev-1.13.0: images and box questions work in the playground and the API, and long_context is an API flag. The moving aliases are excluded because a gateway would drop the headers

TypeSafe also added an opt-in that lets it inspect your queries to find model weaknesses. It states that it still does not train on customer requests.

Pricing and limits

TypeSafe’s published figures and how the same model is billed on Jev AI.

TypeSafe list price
$0.042 per million input tokens, the same as 1.13.0; output tokens are free
Context window
64k tokens per request and 32k for the state plus the longest single question by default; 262k with the long-context header on TypeSafe directly
On Jev AI, playground
1 credit per run, any length
On Jev AI, API
Metered on input tokens: 1 credit per 1,000,000 input tokens, the same rate as jev-latest and jev-1.13.0
Welcome credits
20 on a new account

Rate limits are published per organisation in the TypeSafe console and adjust dynamically. Jev AI plans and credit packs are on the pricing page.

TypeSafe’s internal measurement

Published with the release and labelled by TypeSafe itself "Warning: AI benchmarks. Please don’t trust them." It includes only open-weight rivals, because TypeSafe keeps its validation set private, and TypeSafe notes the numbers are likely biased toward its own models and may not be consistent next time. JevBench v1.7.48 has since measured 1.14.0 separately: #22 of 54 hosted APIs at 70.2, below 1.13.0’s 71.4 at #18.

ModelIntelligence (higher is better)Calibration error (lower is better)
jev 1.14.060.01.6
jev 1.13.058.91.8
kev-27b46.717.8
pplx-decider-v1.1-27b46.035.0
clef44.311.2
clef-flash38.819.1
kev-4b36.910.3
d1-3b33.819.9

TypeSafe’s cost chart places both Jev versions at roughly $130 per million queries, Clef Flash near $150, pplx-decider near $270, Clef near $900 and kev-27b near $3,500. For an external measurement of both versions against hosted and open systems, see the JevBench results page and the Jev 1.13.0 guide.

Call 1.14.0 on Jev AI

Keep jev-latest, which now runs 1.14.0, or name jev-1.14.0 to stay on this version when the alias moves again; jev-preview follows TypeSafe’s newest build. The request below is in Spanish because multilingual decisions are where TypeSafe expects the biggest gain.

curl --fail-with-body https://jev-ai.pro/api/v1/systemone \
  -H "Authorization: Bearer $JEV_AI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-1.14.0",
    "state": "Cliente: quiero cancelar y que me devuelvan el mes que no he usado.",
    "questions": {
      "intent": { "type": "choice", "instructions": "What is the customer asking for?",
                  "criteria": { "cancel": null, "refund": null, "both": null, "other": null } },
      "urgent": { "type": "noul", "instructions": "Does this need a reply today?" }
    }
  }'

Create the key on the Jev API page. The official TypeSafe SDK works against this endpoint with its base URL set to https://jev-ai.pro/api; the documentation has the full request and response reference.

A box question

The same request works against Jev AI with a Jev AI key: send the assets map and Jev AI adds the header. Swap the host and the key to call TypeSafe directly.

curl --fail-with-body https://jev-ai.pro/api/v1/systemone \
  -H "Authorization: Bearer $JEV_AI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-1.14.0",
    "assets": { "photo": { "type": "image", "format": "jpg", "data_base64": "/9j/4AAQ..." } },
    "state": "A shelf photo: `photo`",
    "questions": {
      "mug": { "type": "box", "instructions": "the coffee mug in `photo`" }
    }
  }'
# answer: { "mug": { "type": "box", "box": { "top": 0.25, "bottom": 0.82, "left": 0.44, "right": 0.84 }, "confidence": 0.92 } }

Coordinates are fractions of the image with the origin at the top-left corner, so left < right and top < bottom. The same assets map and header let a photo stand in for text in the state, in a question’s instructions, or as an example image inside the criteria of a choice, score or yes/no question.

Jev 1.14.0 FAQ

What is new in Jev 1.14.0?

TypeSafe describes it as small, subtle improvements across the board rather than a flashy release, with the largest gain expected in multilingual decisions. Its own internal measurement moves intelligence from 58.9 to 60.0 and calibration error from 1.8 to 1.6 against 1.13.0. The visible new capability is the box question type, which only 1.14.0 answers.

Is jev-1.14.0 the same as jev-latest?

Yes, for now. TypeSafe moved jev-latest to jev-1.14.0 after the October 14 release, and jev-preview points to the same build until a newer preview ships. Jev AI sends jev-latest to 1.14.0 on every route. Pin jev-1.13.0 if you want the old answers to stay stable, or jev-1.14.0 if you want to stay on this version when jev-latest moves again.

When is Jev 1.13.0 deprecated?

No date is set. TypeSafe says 1.13.0 stays available for a while and that it intends to give 60 days of warning before formally deprecating it, with 30 days of warning for 1.14.0 when its turn comes. A deprecation policy is to be published first.

Can Jev 1.14 read images?

Yes, as an experimental feature. Put PNG or JPEG files in a top-level assets map and reference each one by key from the state, the instructions or the criteria; TypeSafe needs the header TypeSafe-Danger-Zone-Images: true, which Jev AI sends for you whenever a request carries assets. The same works on 1.13.0. Attach a file in the playground and it becomes an asset with a key you paste into the text. TypeSafe does not call this production-ready, so measure it before you depend on it; Clef, Clef Flash and Jev-Omni remain the non-experimental image options.

What is the box question type?

A question that returns where something is in an image: the smallest rectangle around the target as top, bottom, left and right fractions between 0 and 1, measured from the top-left corner, plus a confidence between 0 and 1. It is the first Jev type that is not a yes/no, choice or score decision, and only jev-1.14.0 answers it, with an image in the same request. Jev AI accepts it in the playground and the API; attach a picture, select jev-1.14.0 and add a Box question naming what to find.

How long can the context be?

By default 64k tokens per request and 32k for the state plus the longest single question, as before. On the Jev AI API, send "long_context": true with jev-1.14.0 or jev-1.13.0 and both limits rise to 262k tokens at the same price per token. TypeSafe kept the default low on purpose: long, noisy state makes every model less accurate, and it recommends filtering before sending.

How much does Jev 1.14 cost?

The same as 1.13.0: $0.042 per million input tokens on TypeSafe with free output tokens. On Jev AI it costs 1 credit per playground run or 1 credit per 1,000,000 input tokens through the API, with 20 welcome credits on a new account.

How does Jev 1.14.0 score on JevBench?

Benchmark Heaven measured it on October 17, 2026 and JevBench v1.7.48 ranks it #22 of 54 on its board for hosted APIs, with a composite of 70.2 against 71.4 for 1.13.0 at #18: intelligence 59.9 against 63.6, calibration 90.1 against 90.6 and speed 91.0 against 91.2. On that board’s Capability ranking 1.14.0 is #17 and 1.13.0 #15. The benchmark keeps 1.13.0 as the reference row that sets its Jev-class caps. TypeSafe's own measurement above covers open-weight rivals only, and TypeSafe itself says not to believe it entirely.

Sources

Read on October 15, 2026; figures can change.

Jev is developed by TypeSafe. Jev AI is a third-party playground and API for Jev models; accounts, keys and balances here belong to Jev AI.