handbooksLimits, pricing, and weak spots
The current Jev model, its context limits, rate limits, and price, plus the short list of things it does badly. One page to check before you ship.
- Current model:
jev-1.13.0. The aliasjev-latestpoints to it. - 64k tokens per request in total. Output tokens are free.
- $0.042 per million input tokens on TypeSafe directly.
- Text only, best in English, 255 options max per Choice, 2 to 10 levels per Score.
- It does not write text, do math, or compare dates well.
The model
| Name | Points to | Notes |
|---|---|---|
jev-latest | jev-1.13.0 | Stable, the SDK default |
jev-preview | jev-1.13.0 | Same model for now |
jev-1.13.0 | itself | Pin this in production so upgrades do not surprise you |
Jev is not fine-tuned per customer. You shape its behavior through state, instructions, and criteria. TypeSafe says it does not train on customer requests, and zero data retention is available for enterprise.
Limits
| Limit | Value |
|---|---|
| Context per request | 64k tokens (state plus all questions) |
| State plus longest question | 32k tokens |
| Choice options | up to 255 |
| Score levels | 2 to 10 |
| Requests | 1,200 per minute |
| Throughput | 250,000 tokens per second |
| Input | Text only: string, JSON object, or array |
Rate limits were being adjusted during launch, so check the models page if you plan for high volume.
Speed and price
- Latency: 70 to 500 ms end to end.
- Price: $0.042 per million input tokens. Output is free.
- TypeSafe's own benchmark: up to 200 times faster and 400 times cheaper than frontier chat models on decision tasks. Treat that as the best case, and measure on your own workload.
The number above is TypeSafe's direct price. Gateways like Vercel AI Gateway, OpenRouter, and Cloudflare set their own. Some third-party guides also list older or different numbers, so check the page for the channel you actually use.
Weak spots
From TypeSafe's own list for jev-1.13:
| Weak spot | What to do instead |
|---|---|
| Reads questions literally | Write the exact condition in instructions |
| Math and counting | Do it in code |
| Dates | Parse and compare in code |
| Double negatives, long chains of logic | Ask a direct, single-step question |
| Long, noisy state | Only send what the question needs |
| Adversarial input | Test with hostile examples, use explicit criteria |
| Instructions that conflict with criteria | Make them say the same thing |
| Opposite questions do not sum to 1 | Ask one version of each question |
| Generating text | Use an LLM, or turn extraction into a bounded Choice |
Where this page's facts come from▾
TypeSafe's models page, state docs, jev-1.13 known issues, and the launch post for the benchmark numbers.