Jev 1.13 bills $42 a billion tokens, and it can’t count

We've got a support inbox that gets triaged by a frontier model at $10 a million tokens, and most of what it does is answer "is this urgent" and "which team". That's a lot of money for a yes or a no. On 15 September 2026 TypeSafe AI, the company Diogo Almeida (one of the RLHF authors behind the original ChatGPT) founded, put a model into early access that only answers questions like those. It's called Jev, the docs list it as jev-1.13.0, and it charges $42 per billion input tokens with nothing at all for output. So we read the launch post, the models page and the part of the docs the launch post doesn't link to, the one titled "jaggedness".

What you get back, and what you never will

Three question types. A Choice picks one option from a set you define, and comes back with a probability for every option plus a confidence number. A Score rates the state against an ordered rubric you wrote (calm, frustrated, very angry) and returns a probability-weighted expectation, so you'll see 1.035 rather than 1. A Noul is a yes/no and returns one number, the probability the answer is yes. You can pack all three into one request, they're evaluated in parallel against the same state, and the docs recommend sending speculative questions you may not need because extra questions don't add latency.

What it can't do is the whole rest of the job. No reply, no summary, no code, no explanation of why. Ask it something that needs a string and there's no field for the string to go in. TypeSafe calls that "zero hallucinations", which is technically true and a bit cheeky: a model that can't emit text can't emit a wrong sentence. The Register made the same point on 16 September. What's actually interesting is the calibration claim, that across many predictions a 0.8 should come true about 80% of the time. That's a property you can test on your own labelled data in an afternoon, and it's the one we'd test before anything else.

The response carries a usage block with input and output tokens, and the output ones are billed at zero. jev-latest and jev-preview both point at jev-1.13.0 today, and the docs say to pin the versioned ID if you've tuned confidence thresholds, because an alias moves when a release ships. There's a Python SDK (pip install typesafe-sdk, Python 3.10 or newer) and a JavaScript one, and a GET /v1/models that lists the aliases.

The price, next to what it replaces

$42 per billion is $0.042 per million. The models page prints both, which we appreciate, because per-billion pricing reads like a marketing unit until you do the division. Against the $10 input rate on Claude Fable 5.1 that's 238x, exactly the figure on TypeSafe's homepage. Against the $2 on GPT-5.6 Terra, the model most triage pipelines we know actually run on, it's about 48x, and that's before you count the output tokens Terra bills at $12 and Jev bills at nothing. TypeSafe's own workflow demo puts one run at $0.000081 and 0.114 seconds against 8.566 seconds for a frontier model.

Now the fine print, which is longer than the price. Input is text only: strings, JSON objects, arrays of text. No images, audio or video, and the docs say "yet". English is the primary training language; other languages, CJK included, are accepted with lower accuracy and a suggestion to watch the confidence field. The request budget is 64K tokens for state plus all questions combined, with a 32K ceiling on the state plus the single longest question. Rate limits are 250,000 tokens a second and 1,200 requests a minute, and there's a warning box saying both "can change without notice" while they let people off the waitlist. There's no fine-tuning: the same weights serve every account, and your domain goes into the state and the criteria. TechCrunch reported the API buckling under demand in launch week, and the docs' own sentence about "upcoming large GPU deals" reads like a company that got more traffic than it planned for. The Register puts the funding at $40 million.

The page you should read before the pricing page

TypeSafe publishes a document called "Jev 1.13 jaggedness", last reviewed 17 September, and honestly it's the most useful thing on the site. Nine failure modes, each with a "do this instead". Jev doesn't count, whether that's characters in a word or items in a list, and the error grows with the size of the thing. It doesn't compare dates; it reads them as text, so "which comes first" is unreliable and gets worse with mixed formats. It's literal: it answers the question you wrote, not the one you meant, so negations and scoping words get read at face value. Double negatives and questions about a property of a property cost accuracy. A large state full of irrelevant detail hurts. Score outputs shouldn't be used to reconstruct a precise number between two rubric levels. And the ninth mode is generation, for which the advice is to use a generative model.

Read that list twice and a pattern falls out. Every fix is the same fix: move the arithmetic into code and leave the model the judgment. Their counting example literally iterates over a list in Python and asks one Noul per item, then sums in code. Their date advice is to extract month, day and year as three Choice questions with an explicit "not stated" option and assemble the date yourself. That's a sound way to build with any model, and it's the opposite of how most teams use a chat model, where the prompt asks for the whole answer and the code trusts the string. I might be wrong, but I think the discipline is the product here more than the weights are.

Which brings us to what we'd do. If you've got a classifier or a router running on a frontier model, this is worth a week. Pull 500 labelled examples and run them through the playground or the API. Then check whether the probabilities are calibrated on your data, not TypeSafe's. If they are, the cost of that pipeline drops by two orders of magnitude and the latency by one. If you've got a guard model already, like the 3B Shieldstral you run locally, Jev competes on the same job at a different price point and with no weights to host. What we wouldn't do is sign an enterprise contract this month. The limits are moving and there's exactly one model version. And a company that says "not yet" about images and non-English text is telling you where 1.14 is going. We'd rather see it land first.

Sources

TypeSafe AI, Introducing System One Models & Jev, Diogo Almeida, 15 September 2026 (the 70 to 500 ms latency, the $0.042 per million input and free output, RLCD, the parallel sampler, early access and waitlist). TypeSafe AI docs, Models, read 19 September 2026 (jev-1.13.0, the $42 per Btok row, the 250,000 tokens per second and 1,200 requests per minute limits, the 64K and 32K budgets, text-only input, the aliases, no fine-tuning, language support). TypeSafe AI docs, Jev 1.13 jaggedness, last reviewed 17 September 2026 (the nine failure modes and the counting and date examples). TypeSafe AI docs, API reference and Confidence, read 19 September 2026 (the request and response shapes, the three primitives, the usage block). TypeSafe AI, homepage, read 19 September 2026 (the 0.114 s against 8.566 s workflow demo, the $0.000081 per workflow, the 238x figure). TechCrunch, A new kind of AI model from a ChatGPT inventor is thrilling developers, 18 September 2026 (Almeida's background, developer cost reports, the launch-week API strain). The Register, TypeSafe AI debuts model for machines that plays Doom, 16 September 2026 (the $40 million funding, the hallucination-free framing critique, the Terra comparison).

Frequently asked questions