Jev AI
TypeSafe's Jev model, run on your own text below with no key. Then its specs, versions and limits, each answered in a paragraph with a source.
Run this example
3 questions · no key neededIndependent and unofficial. Jev is a model by TypeSafe AI; Jagent is not affiliated with or endorsed by them. Facts below are from TypeSafe AI's docs and OpenRouter; sources at the end.
What is Jev AI, in one paragraph?
Jev AI is a decision model built by TypeSafe AI. It reads the text you send — a support ticket, a transcript, a web page — and answers typed questions about that text with probabilities, all in one forward pass. It never writes a sentence back. TypeSafe calls this kind of model a System One model, after the fast, intuitive half of human judgment.
| Made by | TypeSafe AI |
|---|---|
| Released | September 15, 2026 |
| Current model | jev-1.14.0 — since October 14, 2026 |
| Input | Text — images in beta |
| Context | 64k tokens per request |
| Price | $0.042 per 1M input tokens — output is not billed |
| Speed | 70–500ms end-to-end |
Quick answer · verified 2026-09-19
Is Jev AI an LLM?
Not in the sense people usually mean. An LLM such as GPT or Claude reads text and writes text. Jev AI reads text and returns numbers: which option wins, where a passage sits on a scale, or how likely a statement is to be true. Three question types cover nearly everything the Jev AI model does:
- Choice picks one option from a list of up to 255 and returns the winner, a probability for every option and a confidence.
- Score places the text on an ordered rubric you define and returns the probability-weighted level, with a probability for each level.
- Noul answers a yes-or-no statement with the probability of yes, and nothing else.
Jev 1.14 added a fourth, experimental one: Box, which finds a thing in an image and returns its four edges. It is covered under images below, and measured in Jev Box; the other three each have a page too: Choice, Score and Noul.
Because nothing is generated, there is no output to bill and no prose to parse. Why a model class like this exists at all is on the System One page; the exact request and response shapes are in the API reference.
How does Jev AI work?
You send one request holding a state and as many questions as you like, and Jev answers all of them in the same pass — which is why a dozen questions take barely longer than one. The option and level descriptions you write are the whole prompt: rewording them moves the probabilities more than anything else you can change.
TypeSafe trains the model with RLCD, aiming for calibration: when Jev says 0.9, it should be right about nine times in ten. That is what makes the confidence usable — route the clear cases automatically and send the uncertain ones to a person or a bigger model. The same weights serve every customer; there is no per-account fine-tuning or LoRA.

Is Jev AI multimodal?
Partly, and only on request. Text is the default. Since Jev 1.14 (October 14, 2026) TypeSafe accepts images as an experimental input: send the TypeSafe-Danger-Zone-Images: true header with each request that carries one, on 1.13 or 1.14. A PNG or JPEG of 280 to 1,568 pixels a side can sit in the state, an instruction or an option, billed as input tokens at the text rate. The new Box question, 1.14 only, returns where something is in the picture. Audio and video are not accepted. English is where accuracy is best; other languages work, but TypeSafe says not equally well.
What is the Jev AI context window?
64k tokens per request, shared between the state and every question in it. A second limit is tighter: the state plus the single longest question must fit in 32k. Those are the defaults on both 1.13 and 1.14. Sending the TypeSafe-Danger-Zone-Long-Context: true header raises both limits to 262k tokens at the same price per token, which is why OpenRouter now lists typesafe/jev-1.14 with a 262K context.
Fitting is not the same as working well. Long context is slower and less accurate, and accuracy falls as the state fills with text unrelated to the decision, so send the part of a document the question is about. The other limits are on the limitations page.
Is Jev AI free?
No. Input costs $0.042 per 1M input tokens and output is not billed; a decision on this site's fixtures averages 447 input tokens, about $0.0000188. Measured calls, a calculator and other decision models' rates are in Jev AI pricing.
Can you run the Jev model locally?
No. TypeSafe has released no weights and offers no self-hosted version, so the Jev model is not open source. Open models built for the same one-pass decision, such as Plumb-4B, do run locally; on JevBench v1.4.2.1 it scored 65.8 against Jev 1.13.0's 63.3. The full list is on the open source page.
Jev versions: Jev 1.14, Jev 1.13 and jev-latest
Jev shipped on September 15, 2026 as jev-1.13.0. Jev 1.14 followed on October 14, 2026, and the current weights are jev-1.14.0. Two aliases point at them: jev-latest, the newest official release and the SDK default, and jev-preview, the newest build of any kind. Called on 2026-11-04, both aliases answered as 1.14, and jev-1.13.0 still answered as itself.
| Model ID | Released | On OpenRouter | What changed |
|---|---|---|---|
jev-1.14.0 | October 14, 2026 | typesafe/jev-1.14 | Small improvements throughout, plus the experimental Box type. Same price and limits as 1.13; 30 days of notice before deprecation. |
jev-1.13.0 | September 15, 2026 | typesafe/jev-1.13 | The launch model. Still served by its id, with 60 days of notice before deprecation; takes images and long context behind the same headers. |
Aliases move when a release ships, so a confidence threshold tuned on old weights can drift without a line of your code changing — which is what happened to anyone on jev-latest in late October. Once you have tuned, request the versioned ID. Every response names the version that actually answered, which is worth logging next to each decision.
Who makes the Jev model?
TypeSafe AI, which came out of stealth with Jev on September 15, 2026, backed by a $40M seed round led by DCVC. Its co-founder is Diogo Almeida, former OpenAI researcher and RLHF co-inventor. The company sells one product, the Jev API; what it has published and how its claims hold up are on the TypeSafe AI page.
What people build with Jev AI
Anything that ends in a bounded decision rather than a paragraph. Each of these has its own page here, with measured output:
- Use cases
Ticket triage, moderation, routing and nine more, each runnable.
- Jev agents
Tool choice and confidence gates inside an agent loop.
- In Claude Code, Codex and Cursor
The skill, an MCP server and a permission hook.
- Playground
Write your own Choice, Score or Noul question and run it.
- Essay grader
Rubric scores that are model probabilities, not prose.
- Jev MCP server
Yes/no, choice and rubric judgments for any agent.
How to use Jev AI today
- TypeSafe's own API at
https://api.typesafe.ai/v1/systemone, with thetypesafe-sdk(Python) and@typesafe-ai/sdk(JavaScript) packages. Keys come from an early-access waitlist and are issued in batches. - OpenRouter, as
typesafe/jev-1.14; request it by ID, as it is not in the chat model list (details in Jev on OpenRouter). - The Vercel AI Gateway, as
typesafe-ai/jevortypesafe-ai/jev-1.14.0, at the same price. - This site: an API key for an endpoint with the same request shape, and a remote MCP server at
https://jev-agent.com/api/mcpwith seven judgment tools for Claude, Cursor and other agents.
Setup for each route, step by step, is on getting started; the MCP tools are documented on the grading API page.
Jev AI alternatives
Three, depending on the job. A chat LLM when the decision needs reasoning across a long document, or an explanation beside it — Jev vs LLM has the measured trade-off. A fine-tuned classifier such as BERT when the labels never change and the volume is enormous — see Jev vs BERT. And the open models above when the data cannot leave your own servers.
Sources
- TypeSafe AI — current quick start (checked 2026-10-04)
- TypeSafe AI — Score primitive
- TypeSafe AI — confidence routing
- TypeSafe AI — example use cases
- OpenRouter — Jev Router
- TypeSafe AI — official documentation
- TypeSafe AI — Introducing System One Models & Jev
- TypeSafe AI — Jev 1.14: images, Box and longer context (2026-10-14)
- TypeSafe AI — Images (experimental input, limits, token billing)
- TypeSafe AI — Box question type (Jev 1.14)
- OpenRouter — Jev 1.14 pricing & providers
- OpenRouter — Jev 1.13 pricing & providers
- TypeSafe AI — Models reference (versions, aliases, rate limits)
- TypeSafe AI — Jev 1.13 jaggedness (vendor-documented failure modes)
- TypeSafe AI — HTTP API reference
- JevBench v1.4.2.1