# TypeSafe bets against strings > A former OpenAI researcher says the missing piece for automation is a model that returns typed decisions, not text. - Title: TypeSafe bets against strings · Typesafe | Shortify - Summary: A former OpenAI researcher says the missing piece for automation is a model that returns typed decisions, not text. If a model can hand software a typed… - Keywords: ai models, automation, structured outputs, typesafe, latency, technology, startups, bets, against, strings - Source: Typesafe — https://typesafe.ai/blog/introducing-system-one-models-and-jev - Read time: 3 min - Topics: ai models, automation, structured outputs, typesafe, latency, technology, startups ## Superhuman at chat, still no automation Diogo Almeida helped build the instruction-following research behind ChatGPT. Four years on, he kept asking why all that chat ability never turned into real automation. > Models have been superhuman at chat for years, so where is all the automation? ## Two years in stealth TypeSafe spent two years heads-down before shipping anything. Today it announces System One Models, a class built for fast structured decisions software can consume directly. ## A whole new stack Not a fine-tune. New model architecture, a parallel sampler for efficiency, and a training method called Reinforcement Learning for Calibrated Decisions. ## Meet Jev The first public model, in early access now. Similar intelligence to existing LLMs on System One tasks, but roughly two orders of magnitude faster and cheaper. ## It gives up strings Jev cannot generate free text, which is the whole point. No strings means no hallucinations and no type errors, just structured values you defined in advance. > Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out. ## Trained for honest odds RLHF optimizes for what human raters like. RLVR optimizes for what a checker can verify. RLCD optimizes for epistemically honest probabilities on fast decisions. ## Parallel, not one token at a time LLMs generate sequentially, each token conditioned on the last. Jev produces every output in a single hardware-aware query. ## Output tokens are free Input runs $0.042 per million tokens against $0.20 to $10 for LLMs. Output costs nothing, described as too cheap to meter. > Output tokens: FREE (too cheap to meter). ## 70ms instead of 70 seconds Frontier models take 3 to 329 seconds end to end. TypeSafe answers in 70ms to 500ms, which is 40x to 200x faster on System One shaped work. ## Confidence you can act on A model that succeeds 95% of the time but never flags the other 5% cannot automate anything. Jev ships a calibrated probability with every answer. > If a model can do a task 95% of the time but doesn't say when it's in the 5%, it can't automate that task. ## Smart if-statements The target use case is ordinary code with fuzzy branches: classify, route, score, extract, or decide where hand-written rules are too brittle. Surrounding code keeps the model in bounds. ## Map-reduce over petabytes Cheap enough to run across huge datasets for features and insights, fast enough for real-time UX, and cheap enough to sit as a judge or jailbreak detector on every LLM call. ## Receipts, with caveats Speed and price are easy to check. Zero type errors is trivially falsifiable, and the team says it is mathematically impossible to break the schema guarantee. > We love skeptics, and are skeptics ourselves. ## The side-by-side demo A live comparison against GPT-5.6 Terra. They admit the query is simplified, the input short, and the framing favorable, and that the one disagreement looks genuinely ambiguous. ## A new kind of eval Instead of a fixed ground truth, every model runs the same compute graph and is scored against the averaged answers of the biggest models, Astra and Fable. ## Owning the Pareto frontier Across four published workflows, Jev sits nearly two orders of magnitude ahead. The 193.6x faster and 444.6x cheaper headline numbers come from here. > Jev is off the charts. ## Where the bias might be The workflows were built by TypeSafe's own capabilities team, and using OpenAI and Anthropic models as the reference tilts the scoring toward them. ## Type safety is table stakes A hallucinated tool call is annoying inside an agent. Buried three layers deep in a dependency chain with latency guarantees, it breaks the system. > Existing models, no matter how smart, still hallucinate and have type errors. ## Playing Doom and Wikiracing A bot reacting to game state at 10 queries a second costs about $7 an hour. Wikiracing shows what not hallucinating buys you when each step means picking among thousands of links. ## Named after Jevons System One comes from Kahneman's fast thinking. Jev comes from the economist who noticed cheaper coal meant more coal burned, and every 10x drop in the cost of intelligence should unlock 10x more uses. ## Key takeaway If a model can hand software a typed answer with a calibrated probability attached, AI stops being a chat feature and becomes an if-statement you can trust.