What a free output token changes about routing
Jev is $0.042 per million input tokens, output free, 70ms to 500ms. Every-request routing is affordable. TypeSafe's ~68% accuracy means cheap is not correct.
Jev prices input at $0.042 per million tokens, $42 per billion tokens, and charges nothing for output. TypeSafe AI calls the output side "too cheap to meter." End-to-end latency is 70ms to 500ms. Routing, triage, and classification calls that were too expensive to make on every request are, on these numbers, cheap enough to make on all of them.
The primary source is TypeSafe's Introducing System One models and Jev. This page is the cost and latency of routing, triage, and classification on every request. It is not what a System One model is, not the zero-hallucinations claim, and not a typed platform-policy check.
The meter that remains
Jev is not an LLM. It does not generate text. It takes unstructured program state and returns typed structured values plus calibrated probabilities and confidence scores, in a single parallel pass. Output tokens are free. The meter that remains is input, at $0.042 / MTok.
Jev gives up string generation entirely. What comes back is a value inside a declared schema, with a probability and a confidence score. Output cardinality goes up to 255.
Limited early access began on 15 September 2026. The release was announced alongside a $40M seed led by DCVC. TechCrunch is the news report of that launch.
Whose number you are reading
Keep the rows separate when you quote them.
| Figure | Number | Whose number |
|---|---|---|
| Input price | $0.042 / MTok ($42 per billion tokens) | TypeSafe's published price |
| Output price | Free ("too cheap to meter") | TypeSafe's published price |
| End-to-end latency | 70ms to 500ms | TypeSafe |
| Speed versus frontier LLMs | 40x to 200x faster at comparable intelligence | TypeSafe's claim |
| Homepage workflow evals | 193.6x faster, 444.6x cheaper | TypeSafe's homepage figure |
| Bryo AI email classification | 10x to 20x cheaper than Gemini | Nikhil Mudholkar, on the record |
| Vercel safety-command review | 5x to 18x faster, with greater accuracy | Pranit Sharma, on the record |
TypeSafe claims Jev is 40x to 200x faster than frontier LLMs at comparable intelligence. TypeSafe's homepage claims 193.6x faster and 444.6x cheaper on its workflow evals. The homepage pair is TypeSafe's number, a claim about their evals. It is not an independent measurement. "Comparable intelligence" is their phrase. It does not settle the buying decision.
Bryo AI is the deployment with a published cost multiple. Nikhil Mudholkar reports business email classification at 10x to 20x cheaper than Gemini, with better confidence scoring. That multiple is Bryo's reported result against Gemini. It is a different number from the homepage's 444.6x, and a different number from the $0.042 / MTok price.
Vercel is the latency deployment. Pranit Sharma reports that Jev replaced ChatGPT Luna 5.6 for safety-command review, 5x to 18x faster, with greater accuracy. This page uses the speed. "Greater accuracy" is that deployment's report. It does not stand in for TypeSafe's benchmark.
Where the cost case is strongest
The cost case is strongest where the decision set is small and fixed. Routing, on this page, means the call you put in front of an agent to assign a request to a path. Triage and classification are the same economic object. Input is unstructured program state. Output is a typed value plus a calibrated probability and a confidence score, from one parallel pass, with cardinality up to 255.
At $0.042 / MTok in, with output free, the bill is the input you send. At 70ms to 500ms end to end, that pass can sit in front of an agent on every request.
KDnuggets draws the line on novelty: "Classification is not new. Intent detection is not new. Zero-shot classification is not new." TypeSafe optimised an existing shape. It did not invent a new task. The budget change is the price and the latency of a task you could already name.
Jev is not for reasoning, coding, or open-ended generation. It gives up string generation entirely. Jev cannot process images yet. A route that starts from a picture has no line on this price card.
The model class itself is covered on Wikipedia and in DataCamp's account of System One models. Those pages are the nature of the model. This one stops at the bill and the clock.
Cheap on every request can still be the wrong answer
Cheap does not mean correct. TypeSafe's own benchmark shows ~68% accuracy, and the reference answers come from frontier models, not verified ground truth. Independent evaluation is still thin. Run that call on every request and the misses cover the stream. A cheap wrong answer at scale is worse than an expensive right one.
Zero hallucinations means zero out-of-schema outputs, not zero wrong answers. Jev cannot return a value outside its declared schema. It can absolutely still be wrong. This is the single most misread claim about the model. The "mathematically cannot hallucinate" framing is a structural guarantee, not an empirical one. A schema-valid value is what that guarantee covers. Correctness is the ~68% result above.
Armin Ronacher, Earendil CTO, is the rule for the score that comes back with the value. You must interpret the confidence threshold. "if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it." A disregarded result still consumed input. Output tokens were free. Input tokens at $0.042 / MTok were the bill. On every request, the calls you drop are part of the spend, and they are what keeps a 50% label off the path the agent then takes.
Founder Diogo Almeida, ex-OpenAI, who helped build ChatGPT, described the gap as "We have lightning in a bottle, and yet it is not useful." Free output changes the invoice. It leaves that sentence standing. Put the call on every request when the decision set is small and fixed, cardinality stays within 255, the input is not an image, and a 50% probability is discarded. Keep a frontier model where a wrong typed value costs more than the input you would have saved. A full stream of wrong labels is the expensive outcome, at any price per million tokens.