Jev, the AI that writes nothing
Jev picks one of your predefined answers and attaches a probability instead of writing text; a week of outside measurements shows where that is fast and cheap, and where it is not good enough.
TypeSafe AI released Jev in early access on 15 September 2026 with $40M in seed funding led by DCVC; it returns typed decisions with probabilities instead of text, at $0.042 per million input tokens with free output. Its "cannot hallucinate" claim only guarantees the answer's shape, and the game demos (Doom, Minecraft, a drone simulator, a flight search) all read game state written as text, with builder-reported numbers that nobody has reproduced. Independent tests found it level with small OpenAI models on short-text classification at a fraction of the cost, but weak on long documents, arithmetic and multi-step problems, and trustworthy mainly at its confident extremes. For a small game studio it looks like a fast gate inside a loop, not a replacement for rules or reasoning models, and TypeSafe itself says Korean is handled "not equally well".
Jev, released on 15 September 2026, is an AI model that does not write a single line of text. It only picks one of the answers you defined in advance and attaches a probability to it, which makes it fast and cheap. The demos of it playing games got the attention, but a week of outside measurements makes it clear what it is good at and what it is not.
We make games. So when the video of "an AI playing Doom" went around, the first thing we wondered was this. Is it fast enough and cheap enough to actually use inside a game? We gathered a week of material and wrote it up.
What shipped
fact TypeSafe AI, of San Francisco, released its first model, Jev, in early access on 15 September 2026. It came out of stealth the same day, together with a $40 million seed round led by DCVC (TypeSafe launch post, checked 22 September 2026). The waitlist was dropped on 20 September 2026 (UTC) in an announcement from the official account.
fact The company introduces its founder, Diogo Almeida, as a "co-inventor of ChatGPT". What can be verified is that he is the fourth of twenty authors on OpenAI's 2022 InstructGPT paper (arXiv 2203.02155). He puts it more modestly himself: in his own post he says he "took part in building" the method for following instructions.
The name comes from two places. Kahneman's "System 1", meaning fast, intuitive thinking, and the Jevons paradox, in which better efficiency leads to more consumption. The company calls Jev a new class of model, a "System One model".
How it works
fact A request carries the situation written as text or JSON ('state') together with a question whose answer shape is fixed in advance. There are three kinds of question: pick an option, give a score, yes or no. What comes back is only the chosen answer and a probability, with no sentences (TypeSafe docs, checked 22 September 2026).
fact The price is $0.042 per million input tokens, and output is free. Input is text only; images, audio and video are not supported yet. One request takes up to 64,000 tokens, and the state plus the longest question together take up to 32,000 of those (model docs, checked 22 September 2026).
fact The speed claims differ by channel. The launch post says 40 to 200 times faster than existing models, with a response in 70 to 500 ms; the home page says 193.6 times faster and 444.6 times cheaper; the press release says up to 100 times. The company itself wrote that the home page numbers come from its own evaluation and are probably "the high end of real gains" (launch post, checked 22 September 2026).
What "no hallucination" actually means
This was the biggest argument at launch. To put the conclusion first: it guarantees the shape of the answer, nothing more.
fact The company's frequently-asked-questions page says Jev "guarantees the shape of the answer, but does not guarantee that every judgement is correct". Given a list of options it cannot make up an answer outside the list, but it can pick the wrong one (TypeSafe home page, checked 22 September 2026). The CEO also admitted on Hacker News that it "can be confidently wrong".
fact The developer Armin Ronacher told TechCrunch that the model "pushes the hallucination problem onto the user a little". That is, what to do at 50% and what to do at 95% is for the caller to decide (TechCrunch, 18 September 2026).
What the game demos showed
Most of the videos that went around were games. They have one thing in common. The model did not look at the screen. Someone wrote the game state out as text and fed it in.
- fact Doom is a demo TypeSafe built itself. It takes the game state as text data, decides about ten times a second, and by the company's own arithmetic costs about $7 an hour. The company wrote itself that "a scripted bot that is not AI could do better" (launch post).
- fact Minecraft was built by an outside developer. Raising a Canadian flag out of 338 blocks took 2,266 decisions over 16 minutes 31 seconds, and the request round trip measured in an earlier run averaged 128 ms. He wrote himself that this is one success, not evidence that it repeats (repository, checked 22 September 2026).
- fact The drone simulator has ordinary vision code turn the camera feed into JSON, and separate code handles the safety reflexes fifty times a second. Jev only picked a direction about 2.5 times a second: 80 calls over 65 seconds, median 0.11 seconds, and again a single run (repository, checked 22 September 2026).
- fact The web agent company Browser Use finished a Google flight search in 7.1 seconds for $0.0039. It did not go as far as booking, and said the number came from running the same task three times (repository, checked 22 September 2026).
fact In an analysis that collected 12,759 tweets from the four days after launch, the median speed-up users reported themselves was 7 times. That is some distance from the home page's 193.6 (OpenChamber, checked 22 September 2026). Among the game demos we have not yet found one that a third party reproduced with results of their own.
Measured from outside
fact The most detailed result comes from the developer Aman Kumar, who made about 16,000 calls. On short-text classification it matched or beat small OpenAI models on three of four public datasets, at 5 to 56 times lower cost. On banking-enquiry classification into 77 classes it lost, at 76.0%. On work that means reading a whole document it did worse than the small model he uses now, and changing the prompt did not help (his measurement write-up, 18 September 2026).
fact The way he ended up using it is this. Answers with a probability of 0.9 or higher were about 96% correct on the public datasets, and the middle band was close to a coin flip. So trust both ends, and hand the middle to a model that reads documents properly. When he sent 100 at once, a few times in a thousand it took 10 to 35 seconds (same post).
fact Even in the company's own evaluation Jev is not first on accuracy. It averages 67.8% across four workflows, below the 74.1% of the most accurate model. In exchange it costs $0.0004 and 0.4 seconds per item, far cheaper and faster than that model's $0.0836 and 23.3 seconds (TypeSafe evals page, checked 22 September 2026). The company's own docs write the weak spots down directly: arithmetic and counting, comparing dates, reasoning that has to cross several steps, long inputs full of irrelevant material, and hostile phrasing slipped into the input (weak-spot docs, 17 September 2026 revision).
fact The launch post on Hacker News took 1,941 points and 509 comments (21 September 2026, 16:50 UTC, Hacker News). Reactions split hard: not publishing the architecture, not including public benchmarks, and the string of open-source imitations that followed.
Where a small game studio would put it
From here on it is us thinking out loud. Jev is not a tool that designs a game for you. It is closer to a part that takes a situation and quickly picks "which of these should I do now". Seen that way, a few places in games come to mind.
- Short decisions inside the game loop. Whether an NPC runs or fights, which item to pick up — decisions where the options are already fixed. But as the Doom demo shows, writing the game state out as text is our job, and there are plenty of places where a script does better.
- Automated playtesting. Running a button-pressing bot cheaply over many rounds. It will not play as well as a person, but for finding where players get stuck the number of rounds is what matters.
- Short-text judgements. Yes-or-no problems such as filtering nicknames on a leaderboard. Following Aman Kumar's advice, handling only the confident end automatically and passing the rest to another method looks right.
The things that bother us are clear too. The company's docs say that Chinese, Japanese and Korean characters are "handled, but not equally well", and we have not yet found a measurement done in Korean. The servers are on the US west coast, so calling from Korea will be slower than the published latency. The weights are not public, and the default version name changes quietly when a new model arrives. Whether the free output price lasts is something the company itself wrote it "has to prove over the long term".
So our conclusion is this. Good enough as a fast gut feeling, too early to hand the whole judgement to. What rules can do, rules do; what needs long thought goes to a thinking model. The short, frequent choices in between are where Jev belongs.