# Jev: The ChatGPT Co-Creator's System One Model Can't Talk > Source: > Published: 2026-09-17 21:12:47+00:00 Diogo Almeida’s new model cannot talk. Jev, released in early access on 15 September by TypeSafe AI, returns typed decisions with calibrated confidence: a choice from up to 255 options, a score, or a yes/no probability. It is not a language model, and the man who helped build ChatGPT says that is the point. Jev runs 70 to 500 milliseconds end to end, prices input at $0.042 per million tokens, and charges nothing for output. Almeida is the OpenAI researcher behind RLHF and InstructGPT, the methods that became the research behind ChatGPT and GPT-4. What he kept asking after co-inventing it: why have superhuman chat models not led to AGI? Chat is solved. Automation is not. Two years in stealth, $40 million in backing, and a team from OpenAI, Google Brain, Meta and FAIR went into his answer. Jev. The argument underneath it is that chat was the wrong target. A model built to talk is a different thing from a model built to decide. Jev has no chat interface. It cannot generate strings at all. Give it structured state and it returns one of three primitives: a Choice, a Score, or a Noul, each with a confidence value between 0 and 1. Your software acts on the answer. The docs are blunt about the edges: 32K context window, no image input, no code, no prose. The confidence value is the whole design. Calibrated confidence means the number Jev reports matches how often it is right: when it says 0.9, it is right about nine times out of ten. That is not a promise LLM APIs have kept in a form software can act on. Most say 0.9 and mean “sounds plausible”. The “can’t hallucinate” coverage needs a qualifier. Jev structurally cannot emit a malformed answer, the output shape is guaranteed by construction. It can still return a schema-valid answer that is factually wrong, and Almeida said as much during the launch discussion. Two demos tell you more than the benchmark numbers. There’s a video of Jev playing Doom at about ten calls a second, roughly $7 an hour of play. There’s a Wikipedia game where Jev races from page to page choosing among thousands of links. Both make the same point visible: this thing decides fast enough to sit inside a game loop, and cheap enough that nobody notices the bill. The training method is RLCD: Reinforcement Learning for Calibrated Decisions. Almeida’s critique of RLHF is specific. It trained for human preference, which made models superhuman at following instructions, and it also produced overconfidence and mode dropping, the exact failure modes that keep a human in the loop. RLCD trains for the opposite property. A machine can act without asking permission. This matters because trust is the thing that is broken. An API that says 0.9 and is right 60% of the time is not usable in a pipeline, no matter how good the prose is. TypeSafe is betting the market for a decision you can trust is bigger than the market for a sentence you cannot. The headline ranges are 20-200x faster and 40-400x cheaper. They are self-reported, which is worth remembering while the model is waitlist-only. The benchmark video on their proof page compares workflows, not single calls. One example: a System One task that ran in 0.114 seconds at $0.000081 took 8.566 seconds and $0.013880 on a frontier model. Across the workflow suite, the average is 193.6x faster and 444.6x cheaper against GPT-6 Astra and Fable 5.1. The free output is not a promotion stunt, which I assumed at first. There is no autoregressive generation to meter: the architecture returns all outputs at once, in parallel. The comparison models take 3 to 329 seconds per call largely because they generate one token at a time. Jev’s latency is the direct result of skipping that loop. If you build agents, this is the interesting part. Most pipelines use one large chat model as thinker and talker: it classifies the intent, extracts the arguments, routes the work, drafts the response. You pay reasoning cost at every step, even when the step is a decision that fits on one line. In voice pipelines, where I spend my days, the pattern is everywhere. Intent classification, escalation routing, fraud triage. Each is currently a full LLM call with JSON output and a parser you trust about as far as you can throw it. Jev changes the shape of that. Set a confidence threshold at 0.85 and escalate anything below it. That is the escalation pattern production voice agents spend months hand-building, and it stops being a design constraint when decisions cost near nothing. The cleanest framing I’ve seen this week splits the stack: language models for communication, coding models for implementation, decision models for high-volume judgment. Jev is the first decision layer built for that job from the ground up. None of this tells you whether the decisions are any good. Speed and price you can verify in a day. Decision quality takes months of production traffic, and TypeSafe is still a waitlist. The structural claim survives the scepticism though. I have watched whole teams design around the cost of a single LLM choice, caching aggressively and batching calls to cut round trips. A 70ms decision model at near-zero marginal cost is built to remove exactly that behaviour. Almeida has the track record, and early access means the claims get tested before anyone has to commit. If you build automation, the experiment is cheap: put a decision where you used to put an LLM call, and read the confidence score. Jev takes milliseconds, prices input at $0.042 per million tokens, and the waitlist is open.