What is Jev AI?
TypeSafe AI introduced Jev as its first public System One model. It is designed around decisions that software can consume, rather than free-form text generation. The initial release was announced as early access; current account eligibility should be checked in the provider console.
See the launch announcement in the source notes.
Start building with Jev
Model facts, with the gaps left visible
| Detail | Reviewed information |
|---|---|
| Current documented model | jev-1.13.0 · Source |
| Input | Text; JSON object or array containing text · Source |
| Context budget | 64k tokens total; state plus longest question: 32k · Source |
| Question types | Choice · Score · Noul · Source |
| Model parameter count | Not publicly confirmed |
| Model-weight license | Not publicly confirmed |
| OpenRouter availability | Not verified in this edition |
| Independent benchmarks | Not measured by this site |
Use cases to evaluate
- Evaluate a classification task with a defined answer set.
- Route a request to a handler chosen by application code.
- Experiment with a scoring rubric on a labeled evaluation set.
These are starting points for experiments. Test with your own labeled data before relying on a model in production.
Pricing
| Billing unit | Published rate (USD) |
|---|---|
| 1 million input tokens | $0.042 |
| 1 million output tokens | $0 |
Source: Official pricing reference · Checked 2026-09-24.
Open-source status
Not publicly confirmed from the sources reviewed. A public SDK or GitHub organization does not establish a public model-weight release.
Official resources
Jev AI model: typed decisions and chat interfaces
Jev accepts a state and named questions through a decision interface. Your application receives answers associated with those question IDs. The practical distinction from asking for a prose chat response is the contract your code consumes: a selected option, a position on a scale, or a probability for a yes/no condition. Start by identifying the decision your program actually needs, rather than asking for a broad analysis and extracting a label afterward.
This narrower interface is a tradeoff. TypeSafe describes focused judgments as the intended workload; dependent judgments require another request when the next state or question cannot be constructed until the first answer arrives. A single request does not turn its questions into a sequential plan. Keep orchestration in the application, and use the returned values as inputs to explicit code paths.
Official question and composition guide · Reviewed Sep 26, 2026.
Generation is a documented limitation: use a generative model when the required output is newly written text. The provider also warns about numerical precision and literal interpretation. For an integration review, separate language judgments from arithmetic and deterministic validation. This is a division of responsibilities, not evidence that Jev outperforms another model on your workload.
Provider-documented limitations · Reviewed Sep 26, 2026.
Choice, Score and Noul: choose the return shape
Choice for a named destination or class
Choice suits alternatives without an ordering, such as assigning an incoming document to a processing queue. Supply criteria as a map of option names and descriptions. The answer contains type, choice, probabilities and confidence. Your code can route on choice while retaining the full distribution for review. Include an explicit fallback option when real inputs may fall outside the listed classes; otherwise selecting an available label is not proof that the taxonomy covers the input.
Choice request and response reference · Reviewed Sep 26, 2026.
Score for an ordered rubric
Score suits a graded judgment, such as whether a report provides little, partial or sufficient detail. Its criteria are an ordered array, with levels numbered from zero. The answer adds score, legend, probabilities and confidence to its type. A score may sit between levels; it is not necessarily an integer category. Preserve the legend alongside a stored result so a later reader knows which rubric produced the number.
Score levels and response reference · Reviewed Sep 26, 2026.
Noul for a defined yes/no condition
Noul returns type and noul, where noul expresses the probability of yes on a zero-to-one scale. It has no separate confidence field. Optional true and false criteria clarify the condition. A value near the middle means uncertainty about that condition, not a medium amount of a quality. Use a Score when you need degree, and a Noul when you need evidence for a precisely stated proposition.
Noul semantics and response reference · Reviewed Sep 26, 2026.
Context limits and a transparent cost example
The documented Jev 1.13 limit is 64k tokens for the whole request, with a separate 32k ceiling for state plus the longest question. Both constraints matter when packaging several questions. The published price is $0.042 per million input tokens, with no output-token charge.
Illustrative arithmetic, not measured usage: 10,000 requests at 2,000 billed input tokens each total 20 million tokens, or $0.84 at that rate. At 100,000 such requests the calculation is $8.40. These examples assume the stated token count and unchanged rate; they exclude extra attempts and any other services. Count the complete input, including questions, when estimating your own workload.
Official context and pricing reference · Reviewed Sep 26, 2026.
A cheap request is only useful if its decision is reliable enough. Our integration recommendation is to compare total operating cost, including review of mistakes, rather than selecting a model on the token bill alone. The provider warns that irrelevant material can impair judgments. Test compact, relevant states before treating the maximum context allowance as a target document size.
Known limitations with large or irrelevant state · Reviewed Sep 26, 2026.
Five questions before integrating the Jev AI model
- Can this account access the service? The quick start directs users to the console for the Playground and API keys. Confirm your own access before planning a rollout; a public guide does not establish approval for every account. Official quick start.
- Which limits apply to this workload? Check the current tokens-per-second and requests-per-minute limits. The model reference says these can change. Size concurrency from current account conditions rather than assuming the published allowance is a permanent reservation. Official model limits.
- What happens when a request fails? The API documents JSON errors with HTTP status codes: 401 for credentials, 422 for validation, 429 for rate limits and 529 for overload. Verify the actual error payload in your integration; the reference does not establish one universal field schema here. Official error reference.
- Is there a contractual SLA? A guaranteed uptime percentage and support response time for your intended account are Not publicly confirmed by the references reviewed here. Ask the provider for applicable terms; a retry policy is not an SLA. Reviewed API reference.
- When should the application abstain? Determine thresholds with representative labeled examples and the consequences of a wrong action. The confidence documentation recommends handling uncertainty and adjusting thresholds to the domain. No universal threshold is validated by this briefing. Official confidence guide.
References reviewed Sep 26, 2026. Examples have not been run against a paid account.
Jev AI model FAQ
Where is the Jev AI model GitHub repository?
The provider’s GitHub organization lists official Python and JavaScript SDK repositories. Those client libraries are integration resources; their availability does not confirm downloadable Jev weights or a model license. TypeSafe AI on GitHub.
Can I download Jev and run it locally?
A public Jev weight download is Not publicly confirmed in the reviewed resources. Do not mistake SDK source code, an adapter or a fork of another research project for a Jev model release. Reviewed provider repositories.
Where should I start a Jev AI model tutorial?
Use the official quick start to try a state and question in the Playground, then obtain an API key for your own integration. Our step-by-step Jev guide organizes that workflow and marks its testing limits.
How do I sign up for TypeSafe AI?
Follow the console link in the official quick start and use the sign-in or registration options presented there. Instant approval, free credits and eligibility for every region are Not publicly confirmed by that guide.
Does Jev generate a conversational answer?
The documented interface returns typed judgments. TypeSafe lists generation among Jev 1.13’s limitations and directs generative tasks to a generative model. Use that boundary when deciding which part of your application should call Jev. Official limitations.
Can Choice select a class I did not define?
Its answer selects from the supplied options. Define a fallback class when needed, and distinguish that class in your criteria rather than relying on an undocumented response. Official Choice documentation.
Is a Score the probability that something is true?
No. It locates a judgment along your ordered rubric, potentially between levels. The accompanying legend explains the levels. For a yes/no probability, use Noul instead. Official Score documentation.
Does a Noul result of 0.5 mean a neutral rating?
It indicates similar probability for yes and no, not a neutral point on a rating scale. Clarify the condition or choose a graded rubric if neutrality is the concept your application needs. Official Noul documentation.
Can I upload an image or audio file directly?
The model reference documents text input, including text within objects or arrays. Image, audio and video input are not supported there; any conversion into text happens outside that interface. Official input specification.
Does high confidence guarantee a correct decision?
No. Confidence summarizes a distribution, and thresholds depend on the task and the consequences of error. Evaluate against your own examples and retain a review or fallback path. This site has not measured independent accuracy. Official confidence documentation.
FAQ sources reviewed Sep 26, 2026. Unconfirmed items are explicit gaps in these reviewed sources.
Go straight to the source
Reviewed Sep 24, 2026. Provider documentation can change. These are primary sources, not independent performance tests.