TypeSafe's Jev Claims 40-200x Faster Typed Decisions and Free Output

TypeSafe AI opened early access to Jev on September 15, 2026. The official blog prices input at $0.042 per million tokens ($42 per billion) and output at zero. Jev does not generate text.

Founder Diogo Almeida posted the launch at 2:17 p.m. ET as @CompleteSkeptic. The company blog, dated the same day, calls Jev the first public System One model: unstructured state in, typed probabilistic decisions out. Software is the customer. Chat is not.

What TypeSafe says System One is

The September 15 blog defines System One models as frontier models built for fast, structured decisions that code can consume directly. Training is RLCD — Reinforcement Learning for Calibrated Decisions — plus a new architecture and a parallel sampler. Jev is early access, not a general chat API.

Almeida's own line on that page: after two years in stealth, TypeSafe is shipping "a new class of frontier models built to make fast, structured decisions that software can use directly." Think of Jev, he writes, as "a frontier-intelligence function call."

Possible outputs are defined in advance. The blog says the model never makes type errors and attaches calibrated probabilities and confidence to every answer. The Register, September 16, walks the same interface: a state value (JSON or a string like "My card was charged twice") plus question primitives that return structured probabilities. One Register example: {"billing": 0.08, "technical": 0.85, "sales": 0.07} with confidence 0.82.

Speed and cost, source by source

Keep the three number sets apart. They are all company claims.

• Official blog pricing: $0.042 / MTok input, output free. The same table puts typical LLM input at $0.20–$10 / MTok and output at about 5× input.
• Official blog latency: 70ms–500ms end-to-end, versus 3–329 seconds for frontier LLMs. TypeSafe writes that this "can range from 40x-200x faster" on System One-shaped queries.
• Almeida tweet, same afternoon: "20-200x faster" and "40-400x cheaper (w/ output tokens free)." The tweet does not print the $0.042 figure.
• Homepage / workflow evals: 193.6× faster and 444.6× cheaper. The blog says those two numbers come from the workflow evals and "are on the higher end of real world gains." A side-by-side on the site and in The Register: Jev at 0.114s versus GPT-5.6 Terra at 8.566s.

The homepage also claims 238× lower input price than Claude Fable 5.1. The Register repeats that comparison, and quotes GPT-5.6 Terra at $2.00 / MTok input and $12 / MTok output. Those LLM list prices are secondary; the Jev sticker is on TypeSafe's blog.

The architectural claim behind the speed is parallel sampling. LLMs emit one token at a time. System One, TypeSafe says, returns every asked decision in one query.

The trade-off is strings

Jev gives up string generation. The blog is explicit: it is "optimized for structured outputs and can't hallucinate," because it never writes prose, code, or open-ended text. The Register's cut is cleaner: that is not a fair hallucination comparison, because the output is not natural language. A wrong probability is still a wrong answer. It will not invent a legal citation. It also will not write one.

That is the product. Classify, route, score, extract, branch. The surrounding code sets the schema. TypeSafe's listed uses: smart if-statements, map-reduce over large corpora, sub-100ms product paths, and scoring or guardrailing other models. Not chat. Not agents that talk.

$40 million, and how Almeida describes himself

TypeSafe's Business Wire release on September 15, 3:00 p.m. ET, says the company emerged from stealth with $40 million in seed funding led by DCVC. Founders named there: Almeida, Erik Gafni, and Sasha Sheng. Headquarters: San Francisco. Founded 2024. The Register, the next morning, used the same $40 million figure. That number is on the company wire, not Register-only.

Almeida's résumé needs three labels, not one. His launch tweet opens with "After co-inventing ChatGPT." His X bio says he "co-created RLHF/ChatGPT" at OpenAI. The Register calls him a former OpenAI researcher and "one of the co-inventors of reinforcement learning for human feedback (RLHF) and ChatGPT." His own blog is narrower: at OpenAI he "helped build the methods that made language models useful at following instructions and talking with people," and "that work ended up as the research behind ChatGPT."

"I spent years working on models designed to make AI better at interacting with people. But if AI is going to fundamentally change how work gets done, people can't be the only consumers of intelligence."

Diogo Almeida, TypeSafe statement — The Register, September 16, 2026

Doom, then a waitlist

The launch page's favorite demo is Doom: Jev on structured game state, not pixels, at about 10 queries a second. TypeSafe pegs that run at roughly $7/hour. The Register treats the clip as proof the model is for machines. Wikiracing is the other public toy — high-cardinality link choices, cardinality capped at 255, with a two-stage score-then-choose path when the set is larger.

Jev is on a waitlist at typesafe.ai. The names are on the record: System One from Kahneman's fast System 1, Jev from William Stanley Jevons. The bet is that cheaper decisions get used more, not less. Whether that holds is a usage question, not a launch-day proof.

This article was ultrathought.