- Published on
TypeSafe AI Releases Jev: A Decision-Only Model That Returns Typed Probabilities Instead of Text
- Authors
- Name
- DeepHorus
San Francisco AI lab TypeSafe AI has released Jev, the first model in what it calls the System One Models line — a decision-only system that never generates text and instead returns typed, probabilistic decisions that software can act on directly. The launch represents a deliberate break from general-purpose generative models, aiming squarely at classification, routing and scoring workloads where a usable answer matters more than fluent prose. The company was founded by former OpenAI researcher Diogo Almeida together with Erik Gafni and Sasha Sheng. Almeida co-invented reinforcement learning from human feedback and contributed to the research behind ChatGPT, credentials TypeSafe is leaning on to signal technical seriousness.
The motivation behind Jev is economic as much as architectural. Teams that repurposed large language models as classifiers discovered they were paying for token generation and tolerating hundreds of milliseconds of latency simply to produce a label that could have been returned as a number. TypeSafe argues that text generation is overhead for a decision task, and that returning a calibrated probability lets calling code enforce its own thresholds rather than parsing prose. Jev is positioned as the first of a family of such models rather than a one-off release.
Mechanically, a caller sends a state — either a plain string or structured data — alongside a set of typed questions. Jev evaluates every question in a single parallel pass and returns Choice, Score and Noul answers, each carrying a probability distribution and a confidence value, so code can act when confidence clears a threshold and escalate when it does not. TypeSafe prices input at $0.042 per million tokens and makes output free, with a 32,000-token context window and quoted end-to-end latency of 70ms to 500ms. The model is trained using a method the lab calls Reinforcement Learning for Calibrated Decisions, aimed at making confidence values trustworthy rather than merely plausible.
Adoption has been unusually fast. Vercel added Jev to its AI Gateway on day two and reported it reached nearly 13% of paid teams within 24 hours, roughly twice the share captured by the GPT-5.6 family. Netlify followed, LangChain shipped a TypeSafeClassifier integration with model routing plus an AutoMode middleware that screens tool calls before execution, and five independent Elixir clients appeared within days. Vercel engineer Pranit Sharma found a safety classifier running five to 18 times faster than the LLM it replaced.
Reactions have been positive but measured. Bryo AI CTO Nikhil Mudholkar rated Gemini slightly more accurate on email classification while costing 10 to 20 times more, and singled out Jev as the only option returning a real probability. Earendil CTO Armin Ronacher told TechCrunch the design "delegates the hallucination problem a little bit to the user," who must decide whether a 50% probability warrants action, while endorsing model routing as a strong fit. An OpenChamber analysis of 12,759 launch tweets found user-reported speedups at a median of 7x against the 193.6x headline, cost savings at a median of 30x, and latency at a median of 76ms. On Hacker News, developers noted that while Jev cannot emit an invalid type, it can still emit a completely wrong valid value.
TypeSafe is steering adopters toward its jev-1.13 jaggedness page, which documents unreliable counting, arithmetic and date comparison plus accuracy loss on large noisy state, and advises keeping maths in code. The company recommends pinning jev-1.13.0 instead of the moving jev-latest and jev-preview aliases, and offers a System One adapter for benchmarking existing models against the same schema. Further System One Models are expected as TypeSafe builds out what it hopes becomes a standard decision layer beneath generative systems.
