TypeSafe Jev: the First Decision-Only Model Class, Benchmarked and Priced

TypeSafe's own. The table comes from the vendor's published workflow evals, and the guide's caveats are explicit: the workflows were designed by TypeSafe's capabilities team, and the reference answers are averages from GPT-6 Astra and Fable 5.1, which biases toward OpenAI and Anthropic models. Useful as a shape, not as a neutral scoreboard.

Only with the vendor caveat attached. Those are best-row-per-model figures on the same in-house workflow set, and the guide repeats TypeSafe's own warning that its homepage multipliers of 193.6x faster and 444.6x cheaper are the higher end of real-world gains. Your tail latency on your own state size will look worse than 0.4s.

Pin it. The guide notes jev-latest currently resolves to jev-1.13.0 and recommends the versioned id in production, and the alias already moved once during early access. Also note the 12.2x cheaper and 10.0x faster parallel-questions figure is thirteen batched against the same thirteen sent singly, so it is Jev versus Jev, not Jev versus an LLM.

No text at all: the guide describes a decision-only model returning choices, scores and booleans, so any prose still needs a generative model. And Choice caps at 255 options, which rules out open-ended ranking; you would enumerate candidates yourself and compose the decision in code, as the primitives are meant to be.

100,000 tokens per second and 80 requests per second, with a warning that the limits are adjusting dynamically and can change without notice while TypeSafe adds capacity; exceeding either returns 429, which the official SDKs retry with backoff.

64k tokens per request, with 32k of that for the state plus the longest question; Cloudflare Workers AI and Vercel AI Gateway each list a 32,000-token context window for their hosted Jev.

No. Requests and responses are not used for training, zero data retention is available on enterprise terms, and Jev is not fine-tuned or LoRA-adapted on customer data; every account gets the same weights.