Jev, the model that does not talk: it picks between options up to 200 times faster and 400 times cheaper
Diogo Almeida, co-creator of ChatGPT, left OpenAI and spent two years in stealth mode. His company, TypeSafe AI, now presents Jev. The model comes from a new training method called RLCD. According to its numbers, it is 20 to 200 times faster and 40 to 400 times cheaper than frontier models.
Jev is not an LLM and it does not generate text. It receives a set of options and returns the probability of each one. In the invoice demo, an LLM takes 8.5 seconds to write its answer. Jev picks between fraud, clean or manual review and says it is clean with 88% confidence.
The most viral demo shows Jev playing Doom in real time, at about 10 decisions per second. Rafa explains the trick: Jev does not see the screen. An MCP sends it the game state and the possible actions, and Jev picks one. Its tool call error rate is the lowest in the comparison, but it compares against a different type of model.
Rafa sees a clear use case: classify if a command is safe before it runs, like the Claude Code auto mode, but in OpenCode or other harnesses. It also works for browser use and computer use, if you expose the options as a simple API. Rafa got access last night and promises to share his experience. Watch in the video.
All the news we covered