Jev's Price and Speed Claims, Checked

What TypeSafe published

The launch post of September 15, 2026 gives a latency range of 70 to 500 milliseconds against 3 to 329 seconds for frontier language models, which is where the 40 to 200 times figure comes from. It adds 193.6 times faster and 444.6 times cheaper on production workflows, and zero hallucinations with zero type errors.

The pricing is unusual enough to be worth stating twice: input is metered by the billion rather than the million, and output is free. TechCrunch's coverage on September 18 reported customer figures alongside it — Vercel measuring 5 to 18 times faster, another team 10 to 20 times cheaper than Gemini.

Every one of those numbers comes from TypeSafe or from its early users. That does not make any of them wrong. It makes all of them unverified.

A claimed range, and the one measured average

LangChain's experiment of September 20 reports an average of 0.44 seconds per call. Put that on the same axis as the launch post's range and it lands inside it, near the slow end: 440 milliseconds against a claimed 70 to 500.

Both things are true at once, and that is the useful observation. The vendor's range is not contradicted. The headline multiplier is not reproduced either, because a multiplier is a ratio and the baseline does the work. Against a frontier model that takes three seconds, 440 milliseconds is roughly seven times faster, not two hundred.

What the same run cost

The cost comparison is less ambiguous. In LangChain's experiment a Jev judgement cost $0.00035, and the complete run of repeated judgements came to $0.34. The same run with Claude Sonnet 4.6 came to $28.17, a difference of about 83 times.

Eighty-three is not four hundred and forty-four. It is also more than enough to change what a team does: at a third of a dollar per evaluation pass, you stop choosing between coverage and budget and start running the judge on every trace instead of a sample.

Read the caveats with the numbers

LangChain wrote its own limitations down, which is the mark of a test worth reading. The experiment is built on five weather requests replayed through one agent. The variance result is described as observational, not as evidence that the training objective caused it. And they warn that low cost can amplify mistakes, because a consistently wrong evaluator produces bad feedback at scale.

They also pinned the versions and published the repository, so the run can be repeated. That is the difference between a benchmark and a boast, and it is rarer than it should be five days after a launch.

Why free output tokens matter more than the multiplier

A System One model returns a typed value, not a paragraph, so its output is tiny by construction. Charging nothing for it is less a discount than an admission that there is almost nothing to charge for. The meaningful number is the input price, which is where a long trace lands.

TechCrunch reached for the Jevons paradox to describe what follows, and the analogy fits the evaluation case exactly: when a judgement becomes cheap enough to stop counting, teams do not spend less. They judge far more, on more dimensions, more often. The budget moves from deciding whether to evaluate to deciding what to do with all the feedback.

How to check the next launch

Find the primary source and its date before repeating a number. Ask who ran the test and what relationship they have with the vendor. Look for the method: how many repetitions, against which baseline, on what task. Then ask whether the task in the test resembles the task you actually run, because a model can be excellent at scoring weather answers and untested at everything else.

Finally, check availability. Jev has no open weights, sits behind an early-access waitlist, and reaches most people through infrastructure such as Cloudflare's AI Gateway. A number you cannot act on is trivia.

• System One models: what Jev actually does
• Jev in an agent loop: the decision layer