New model dropped and it is not another chatbot. It is called Jev, from TypeSafe AI. They came out of stealth Sept 15 with a $40M seed, founded by Diogo Almeida, who worked on the RLHF and InstructGPT research that became ChatGPT. The whole idea: it does not generate text. No sentences, no code, no explaining itself. You hand it your program state plus typed questions, and it hands back decisions with calibrated confidence numbers your code can act on directly. They call this a System One model, the fast reflex layer, instead of a model that thinks out loud. Why the numbers are wild: - 70 to 500ms responses, vs 3 to 329 seconds for normal frontier models on the same workflows - $0.042 per million input tokens. Output is not metered at all, because it never generates tokens one at a time - On their own workflow evals: 193.6x faster and 444.6x cheaper than an LLM doing the identical job - 0% type errors by design, not by benchmark. Schema matching is guaranteed, so it cannot hand your code garbage JSON - Trained with RLCD (reinforcement learning for calibrated decisions), so the confidence score actually tracks how often it is right - They demoed it playing Doom in real time at 10 queries per second for about $7 an hour How I would actually use it: keep Claude and Codex for building. Use Jev for the decision layer inside what you build. Routing a support inbox, classifying, "does this need a human", picking 1 of 200 options. You set a confidence threshold. Above it, your code acts on its own. Below it, a human reviews. That is exactly the part where LLMs get expensive and randomly break your JSON at 2am. You can try it today. It is live on Vercel AI Gateway as typesafe-ai/jev with a 32K context, currently listed free, and you call it with the AI SDK experimental_evaluate function instead of a chat completion. Direct early access from TypeSafe is still a waitlist. Being honest about the caveats: - Every benchmark so far is TypeSafe's own, on evals they designed. No independent numbers yet