Published
Updated
TypeSafe leaves stealth and ships Jev, a model that refuses to chat
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
Former OpenAI researcher Diogo Almeida announced Jev, a System One model built for typed software decisions rather than text generation, following two years in stealth. On September 30, 2026 this desk scored 32 labeled tickets. jev-1.13.0 and gpt-6-luna were each right on 32 of 32. p50 wall time was 913.5 ms and 2,698.5 ms.
On September 15, 2026, Diogo Almeida, formerly involved with RLHF work at OpenAI and currently CEO of TypeSafe AI, published a launch thread announcing Jev.
In the announcement, Almeida argued that conversational chat interfaces create bottlenecks for software automation. After two years in stealth, TypeSafe developed Jev to serve as a programmable function call rather than a chat interface.
Jev is TypeSafe’s first System One model. The API accepts unstructured state, such as text or program data, alongside structured questions: categorical choices, numeric scores, or boolean values. The model returns corresponding selections with calibrated probabilities rather than generating text or JSON strings.
Headline metrics cited in TypeSafe’s launch materials include:
- roughly 20 to 200x faster execution than comparable language model workflows
- roughly 40 to 400x lower cost, with output tokens billed at zero
- calibration optimized for decision confidence rather than prose generation
These metrics represent TypeSafe’s internal benchmarks. The announcement post includes workflow evaluations and methodology notes, adding that the top figures of 193.6x speed and 444.6x cost reduction likely represent the high end of real-world performance.
The name references economist William Stanley Jevons and the Jevons paradox, which posits that lowering resource costs increases total consumption. TypeSafe anticipates that low-latency, low-cost decision calls will expand overall automated usage.
On September 30, 2026 this desk scored 32 labeled tickets, eight each in billing, sales, spam, and technical, from an Intel Core i3-12100. The alias in the request was jev-latest. The response model was jev-1.13.0. Jev and gpt-6-luna were each right on all 32. p50 wall time was 913.5 ms for Jev and 2,698.5 ms for Luna. p95 was 1,412.9 ms and 5,227.1 ms. Jev used 12,597 input tokens. At $0.042 per million input tokens on the docs, and $0 for output, the list cost is $0.000529. The Luna gateway returned 1,689 input tokens and 160 output tokens and no dollar amount. A keyword function on the same text scored 30 of 32 and missed two spam lines. One Jev call that answered eight questions about a single email took 895 ms. Eight serial Luna calls for those questions summed to 30.1 seconds, and three of the eight replies wrote a sentence instead of the requested label. Luna wrote a two-sentence apology naming invoice 4412. Jev has no question type that returns that sentence. The 193.6 times and 444.6 times figures stay the company’s high end on four workflows this run did not include.
Access at launch required joining a waitlist, accompanied by a Discord community for developers. TypeSafe posted an initial greeting on X the same afternoon.
This digest draws on that launch thread as the primary public announcement. Subsequent stories document early builder implementations.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Almeida's launch thread, TypeSafe's announcement post, and the company blog align on the launch date (September 15, 2026), architecture (typed decisions without generated strings), and target metrics (20 to 200x faster, 40 to 400x cheaper, with free output tokens).
The 193.6x speed and 444.6x cost reduction claims reflect TypeSafe's internal workflow benchmarks, which the company notes represent the high end of typical gains. We did not reproduce these tests independently.
On September 30, 2026 this desk ran a different set from an Intel Core i3-12100. Thirty-two support tickets, eight each for billing, sales, spam, and technical, went out one HTTP call at a time. The request asked for jev-latest. Every response named jev-1.13.0. Jev was right on 32 of 32. p50 was 913.5 ms and p95 was 1,412.9 ms, including the round trip. The calls used 12,597 input tokens and 1,440 output tokens. At the docs list price of $0.042 per million input tokens, with output at $0, that is $0.000529. gpt-6-luna, on a Responses gateway with reasoning effort set to none, was also right on 32 of 32. p50 was 2,698.5 ms and p95 was 5,227.1 ms, on 1,689 input tokens and 160 output tokens. The gateway returned no dollar field, so this note has no Luna price. A local keyword function scored 30 of 32, at no wait and no cost. It missed two spam lines both models caught: a redelivery-fee message, and a work-from-home offer.
One email with eight questions in a single Jev call took 895 ms, 640 input tokens, $0.000027 at list price. The eight answers were a queue, a language, a priority, a frustration score, and four yes/no values. One Luna call asked for the same eight fields as JSON and returned in 2,963 ms, 104 input tokens and 41 output tokens. Eight serial Luna calls, one question each, summed to 30.1 seconds. Three of those eight replies left the requested label and wrote a sentence about the refund. A two-sentence apology that had to name invoice 4412 is a writing task. The endpoint has no string-generation question, so that apology stayed on Luna, which wrote it in 3,577 ms.
The multiple near 34 times is those eight serial calls against the one Jev call. The multiple near 3 times is one packed Luna call against one Jev call, on p50 wall time from this host. Jev also spent more input tokens than Luna on the 32 tickets, because each call carried the rubric. The 20 to 200 times and 40 to 400 times ranges, and the 193.6 times and 444.6 times workflow figures, remain TypeSafe's measurements on four workflows this desk did not rerun. Published times of 70 to 500 ms are the company's figure. The times above are full HTTP time. TypeSafe's Master Customer Agreement section 2.3(f) forbids customers from publishing benchmarks of the Services. The figures in this note are this desk's measurements, with the version, the date, and the sample above. No invoice was downloaded.
Compare
Subsequent builder demonstrations (such as Browser Use, Stagehand, and Tetris) show millisecond-to-second response times and sub-cent costs on narrow workflows. Those numbers are author-reported demo measurements rather than third-party lab evaluations.
Partner listings on Vercel, OpenRouter, Cloudflare, Netlify, and Venice replicate TypeSafe's documented ranges. The September 30 measurement in Verify is a 32-ticket desk set. It leaves the company's four workflow evals, and the 193.6 times figure, where the company put them.
Terms
- System One
- TypeSafe's term for a model returning structured decisions with probabilities rather than text output, inspired by Daniel Kahneman's distinction between fast and slow cognition.
- RLCD
- Reinforcement Learning for Calibrated Decisions, TypeSafe's training methodology intended to align model confidence scores with true outcome probabilities.
- Jevons paradox
- The economic observation that increases in resource efficiency can expand total demand. The product is named around this economic hypothesis.