The cheapest model released last week refuses to write a sentence.
TypeSafe AI shipped Jev on September 15. It generates no text at all. You hand it program state and typed questions, and it returns a choice, a score, or a probability, all in one parallel pass with a calibrated confidence attached. Input runs $0.042 per million tokens. Output is free. Responses land in 70 to 500 milliseconds.
The argument underneath it is the part worth sitting with. An agent does not only write the sentence your customer reads. Before that it classifies intent, picks a tool, decides whether a command is risky, scores urgency, and routes the ticket. Every one of those has a small, knowable answer set. Run them through a frontier chat model and you are paying a writer to do a clerk’s job, then parsing the writer’s prose to find the answer.
TypeSafe’s own evaluations put Jev at up to 193.6x faster and 444.6x cheaper on narrow decision tasks. Those are vendor numbers and the company says so, flagging that its team built the workflows and that the competing models ran through TypeSafe’s own adapter. The harder signal is distribution. Vercel added it to the AI Gateway on September 16 and developers moved fast.
Founder Diogo Almeida is ex-OpenAI and worked on the research behind ChatGPT. He named the model after William Stanley Jevons, whose observation was that making coal engines more efficient increased coal consumption. That is the bet stated in the branding.
The question for anyone running agents in production: what share of your model calls are decisions rather than generation, and what would those cost at classifier prices?
The rest of the week, including what Anthropic and California each did about who watches these systems, is in the full issue.
https://www.teamignite.vc/blog/last-week-ignite-september-20-2026-the-decisions-got-cheap-the-oversight-got-funded