Jev drops the waitlist — every new user gets 120M free tokens
The decision model everyone has been arguing about is suddenly free to try, and the biopharma world-model crowd just got another $20 million. Also today: the case for paying for calibrated confidence.
TypeSafe AI opened Jev to everyone on Sunday, scrapping the waitlist and giving every new account $5 in credit — roughly 120 million input tokens at the model's published price. The company announced the change on X and pointed users straight at its console; 36Kr reported the credit figure the same day. It lands eleven days after a launch where demand outran capacity badly enough that the API briefly stopped responding, so this is less a marketing push than a lab deciding it can finally serve the traffic it attracted.
The pricing is the reason the number is so large. Jev charges $0.042 per million input tokens and does not bill output at all — the model returns a typed answer instead of generating a paragraph, so there is nothing to meter. That is the entire thesis behind TypeSafe's "System One" pitch: you give it a program state and a set of pre-defined questions, and it comes back with a Choice from a list of up to 255 options, a Score on a scale you set, or a Noul — a probability that a claim is true — each carrying a confidence estimate. No JSON prompt engineering, no parser, no Markdown-wrapped surprise.
The launch numbers TypeSafe published claim 20 to 200 times the speed of comparable LLM calls and 40 to 400 times the cost advantage, with end-to-end latency between 70 and 500 milliseconds; the company's own workflow evals put the gap at up to 193.6 times faster and 444.6 times cheaper. Third parties have started producing their own results. Vercel's AI Gateway lists Jev as its fastest-adopted model ever, and Netlify added it to its gateway in mid-September with no configuration required.
The more interesting signal is what people do with a classifier that tells you how sure it is. Armin Ronacher, CTO at Earendil, framed it as a transfer of responsibility: the user decides that a 50% probability is ignorable while a 95% one is actionable. Vercel engineer Pranit Sharma said swapping Jev in for GPT-5.6 Luna on a command-safety classifier made it five to 18 times faster with better accuracy. It is not a clean sweep — Bryo AI's Nikhil Mudholkar found Gemini slightly more accurate on classifying business email, at 10 to 20 times the cost — and Jev cannot chat, write, or reason in prose. The bet is that most automation never needed any of that.
Mithrl raised a $20 million Series A led by Obvious Ventures to push its biomedical world model deeper into pharmaceutical R&D systems, with Headline, AGI House and several pharma executives joining the round. The company's Mithrl-1 model is positioned as a middle layer built from curated peer-reviewed science rather than extrapolated from public data, deployed privately inside customers and answering natural-language questions, generating target hypotheses and citing where each conclusion came from. CEO Vivek Adarsh told Axios the money funds a second generation of the platform. It is the same wager Novo Nordisk made when it handed Anthropic its drug discovery pipeline — pharma wants a model that can be audited, not one that sounds confident.
Does a probability you can set a threshold on beat a fluent paragraph — or is calibration just the next thing we learn to ignore? Tell us in the comments.
