Jev

2026-09-15 · Publisher : TypeSafe AI · Domain : Artificial Intelligence

TypeSafe AI, a San Francisco start-up, launches Jev in limited early access: a “System One” model that returns typed decisions with probabilities rather than text.

### How it differs from an LLM - Output: an LLM generates text token by token. Jev, according to the vendor, samples typed answers (choice, score, yes/no) defined in advance in parallel, each accompanied by probabilities and a confidence score. - Use: Jev produces no text, no code and no explanation. The documentation presents it as a decision function to plug into code, not as a replacement for the model of a coding agent. - Input: text only (string, JSON), no image, audio or video for now. Jev is a specialized model, alongside LLMs and not in their place. ### Specifications - Announcement: September 15, 2026, limited early access (waitlist, console login). The vendor also says it briefly lost the ability to serve the API because of demand (TechCrunch). - Model `jev-1.13.0` (alias `jev-latest`), endpoint `POST /v1/systemone`. Context of 64k tokens per request (32k for the state and the longest question), 250,000 tokens per second and 1,200 requests per minute (limits adjusted dynamically). - Three question types: Choice (up to 255 options), Score (ordered scale) and Noul (yes/no probability). Python and JavaScript SDKs. - Training: in-house RLCD method (Reinforcement Learning for Calibrated Decisions), on synthetic data according to TechCrunch. - Founders: Diogo Almeida (CEO), Erik Gafni, Sasha Sheng. 40 M USD seed round led by DCVC. - Use cases cited: routing and classification, guardrails and verification of LLM outputs, agent tool calls, real-time applications, a demo playing Doom. ### Pricing | Item | Price (USD) | |---|---| | Input tokens | 0.042 per million (42 per billion) | | Output tokens | Free | Cache and batch: not disclosed. The price is charged on input only. The vendor acknowledges it cannot prove that this price is not subsidized. ### Benchmarks No standard academic benchmark is published. The vendor highlights its own “workflow evals” (scores published by the vendor), with the average of GPT-6 Astra and Fable 5.1 as the reference. | Measure (vendor claim) | Value | |---|---| | End-to-end response time | 70 to 500 ms | | Speed versus frontier LLMs | 40x to 200x (maximum 193.6x) | | Cost versus LLMs | maximum 444.6x cheaper | TypeSafe specifies that these figures are at the high end of real gains, that the evaluations were designed by its team and that the reference favors OpenAI and Anthropic models. ### Vendor claims and caveats Vendor claims, not independently verified: “40x to 200x faster” than frontier LLMs, and Jev “cannot hallucinate” or make a type error. For the vendor, the absence of type errors follows from the closed output format. Independent analyses note that this does not guarantee that the chosen value is the right one. The documentation itself admits that Jev reads instructions literally, counts poorly, lacks numerical precision and does not generate text. Calibration is assessed on groups of predictions, not on a single answer. ### Key takeaways Jev targets fast, cheap decisions inside software (routing, scoring, guardrails), as a complement to an LLM. The announced gains rest on the vendor's evaluations: test them on your own cases before any decision.

Tags : IA, Jev, TypeSafe AI, System One, Décisions typées