Jev
Jev is a model for software, not for chat. You give it some state (text or JSON) and one or more typed questions. It returns a pick from a list, a score on a scale, or a yes/no probability, plus how confident it is. It never writes sentences. TypeSafe AI announced it on 15 September 2026 as the first of what it calls "System One models".
The pitch is speed and price. TypeSafe says Jev answers in 70–500 milliseconds and costs $0.042 per million input tokens, with output free. On TypeSafe's own four-workflow test it scored 67.8% accuracy at about $0.0004 per case, against $0.03–$0.18 per case and 10–38 seconds for the LLMs it compared. These are vendor numbers; nobody has verified them independently yet. Jev is in early access through a waitlist.
Jev is a promising, narrow tool: a very cheap, very fast way to make yes/no, pick-one and score decisions inside software. It is not a chatbot and not a general AI model.
What stands out:
- Price. $0.042 per million input tokens, with output free, is far below even the cheapest mainstream LLMs.
- Speed. A claimed 70–500 ms per call suits real-time routing, where a 10-second LLM call is too slow.
- No parsing. Answers always come back in the shape you asked for, so there is no broken JSON to handle.
The catches are real. Every performance number so far comes from TypeSafe. On its own test Jev roughly matched one frontier model's accuracy (67.8%) but trailed the best one it compared. It cannot do maths, compare dates reliably or explain why it chose an answer, which matters for audits. It is text-only and waitlist-only.
Pick it if you run thousands or millions of classification, routing or triage decisions a day and can test it on your own data.
Skip it if you need written answers, reasoning you can read, image input, or a proven production track record. Use a small LLM such as Gemini 3.5 Flash or Claude Haiku 4.5 instead.
Score breakdown
Best for
- Routing support tickets or agent steps to the right handler
- High-volume classification and tagging
- Scoring leads, content or risk against a fixed rubric
- Yes/no guard checks inside AI agents
What we like
- Extremely low cost: $0.042 per 1M input tokens, output free
- Claimed 70–500 ms responses, fast enough for real-time routing
- Always returns the requested type, so no output parsing or format errors
- Returns probabilities and confidence, useful for setting thresholds
- Answers many questions in parallel in a single call
Watch out for
- Cannot write text, code or explanations
- All benchmarks so far are TypeSafe's own
- Weak at arithmetic, counting, dates and double negatives (per TypeSafe's guides)
- Waitlist-only early access; text input only
Specs at a glance
| Developer | TypeSafe AI (San Francisco) |
|---|---|
| Model type | System One model: returns typed, probabilistic decisions, no text generation |
| Question types | Choice (pick from a list), Score (rate on a scale), Noul (yes/no probability from 0 to 1) |
| Choice limit | Up to 255 options per question |
| Score levels | 2 to 10 levels |
| Context window | About 64,000 tokens shared by state and questions; about 32,000 tokens per single question |
| Input | Text only (strings, JSON objects or arrays); no images, audio or video |
| Output | Typed answer plus probability distribution and confidence |
| Training method | Reinforcement Learning for Calibrated Decisions (RLCD), per TypeSafe |
| API | POST https://api.typesafe.ai/v1/systemone, model route jev-latest |
| SDKs | Python (typesafe-sdk), JavaScript (@typesafe-ai/sdk), Vercel AI SDK |
| Rate limits (early access) | 250,000 tokens per second, 1,200 requests per minute (as reported) |
| Weights | Closed; API only |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| TypeSafe four-workflow eval: accuracy | 67.8% | TypeSafe (via DataCamp) | Vendor benchmark; compared LLMs scored about 68–73% |
| TypeSafe four-workflow eval: cost per case | About $0.0004 | TypeSafe (via DataCamp) | Compared LLMs: about $0.03–$0.18 per case |
| TypeSafe four-workflow eval: latency | About 0.4 s | TypeSafe (via DataCamp) | Compared LLMs: about 10–38 s |
| Speed and cost vs compared LLMs | 193.6x faster, 444.6x cheaper | TypeSafe launch post | Vendor claim, not independently verified |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| Input tokens | $0.042 per 1M tokens | $42 per billion tokens |
| Output tokens | Free | Not metered |
| Access | Early access | Waitlist at console.typesafe.ai; no public free-credit programme announced |
How Jev works, in plain terms
Think of Jev as a smart if statement. Your code sends two things:
- State: whatever the decision is about, such as an email, a support ticket or a JSON record.
- Questions: each one typed. For example, "Which team should handle this? (billing / tech / sales)", "How urgent is it, 1–5?" and "Is the customer angry?"
Jev answers all the questions in one pass and returns, for each, the answer, a probability for every option and a confidence value. Because it never generates free text, it cannot drift into a paragraph or produce invalid JSON. TypeSafe also claims this design means it cannot hallucinate in the usual sense, though it can still pick the wrong option.
The three question types
| Type | What you ask | What you get back |
|---|---|---|
| Choice | Pick one of up to 255 options | Chosen option, probability per option, confidence |
| Score | Rate on a scale of 2–10 levels | Score, probability per level, weighted mean |
| Noul | Is this statement true? | A probability from 0 to 1 |
For more than 255 options, TypeSafe suggests a two-stage approach: first pick a group, then pick within it.
Jev vs using a small LLM
| Jev | Typical small LLM API | |
|---|---|---|
| Output | Typed decision + probabilities | Free text (you parse it) |
| Speed | 70–500 ms (claimed) | Often 1–10+ s with reasoning |
| Price | $0.042/1M input, output free | Usually $0.10–$1+ per 1M input, output charged |
| Explains itself | No | Yes, in words |
| Maths, dates | Weak | Better |
| Images | No | Often yes |
| Access | Waitlist | Open sign-up |
Jev is worth testing when the decision set is fixed and volume is high. When the task is open-ended or needs an explanation, an LLM remains the better tool.
Known weaknesses
TypeSafe's own guides and early reviewers list clear limits:
- Not a calculator. Arithmetic and counting are unreliable.
- Dates and comparisons. It struggles to compare dates or similar values.
- Tricky wording. Double negatives and indirect questions cause errors.
- Noise. Large amounts of irrelevant context hurt accuracy.
- No rationale. It gives probabilities but no written reason, which reviewers flag as a problem for debugging and regulated audits.
Pre-compute anything numeric in normal code and pass Jev a clean, focused state.
- Do arithmetic, date maths and counting in your own code first, then pass Jev the result. Its guides say it is weak at these.
- Use the returned probabilities: auto-act above a high threshold (say 0.9) and send low-confidence cases to a human or an LLM.
- Bundle all questions about one record into a single call. Jev answers them in parallel, so you save round trips.
- Keep the state short and relevant. Irrelevant context lowers accuracy.
- Run a side-by-side test on a few hundred of your own labelled examples before trusting the vendor's 67.8% figure.
Jargon explained
- Typed decision
- An answer in a fixed format, like one option from a list or a number on a scale, that a program can use directly.
- Calibration
- How well a model's confidence matches reality. A well-calibrated model that says 80% is right about 80% of the time.
- Latency
- How long you wait between sending a request and getting the answer.
- Token
- A chunk of text a model reads, roughly three-quarters of an English word.
Alternatives to consider
Frequently asked questions
What is Jev?
Jev is TypeSafe AI's first "System One model". It returns typed decisions (a choice, a score or a yes/no probability) instead of writing text. It is built for software and AI agents, not for chatting.
How much does Jev cost?
$0.042 per million input tokens. Output tokens are free. Access is through early access with a waitlist as of 23 September 2026.
Is Jev better than GPT or Claude?
Only for narrow decision tasks, and only on TypeSafe's own tests so far. There it was far faster and cheaper, with accuracy close to some frontier models but below the best. It cannot do general tasks like writing or coding.
What is Jev's context window?
About 64,000 tokens, shared between the state and your questions, with roughly 32,000 tokens for a single question.
Why is it called Jev?
It refers to Jevons Paradox: when something becomes much cheaper, people use far more of it. TypeSafe expects cheap decisions to be used everywhere in software.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Introducing System One Models & Jev (TypeSafe AI)
- TypeSafe AI documentation: Introduction (TypeSafe AI)
- Jev: TypeSafe's System One Model (DataCamp)
- A deep dive into Jev, TypeSafe's System One model (Flavio Copes)
- TypeSafe AI exits stealth with $40M to build AI for use by software (SiliconANGLE)
- TypeSafe AI's Jev offers an alternative to LLMs that claims to be 193x faster and 445x cheaper (Tom's Hardware)