Jev pricing: TypeSafe AI's structured-decision model costs $0.042 per million input tokens with output free - the arithmetic for a Pakistani software house, the claims labelled, and what to do on Monday
$0.042 per million input tokens. Output tokens free. That is the published price of Jev, the model TypeSafe AI put into early access on 15 September, and it is the only number in the launch that a buyer can check today without trusting the vendor. Everything else in this piece is either arithmetic on that price or a claim with a label on it.
What Jev is, in the vendor's words: a "System One" model that returns typed, schema-constrained values with a confidence score, instead of text. It does not generate strings, so it cannot return a paragraph where you asked for a category. Diogo Almeida, the founder, was at OpenAI on the work that became ChatGPT and has spent two years on this. The post went to the top of Hacker News and had passed 1,100 points and 330 comments by Wednesday morning.
The arithmetic
Take an RFQ triage job: one document in, one of a dozen categories out, with a confidence figure. Call it 2,000 input tokens per document. At $0.042 per million that is $0.000084 a document, or $8.40 for a hundred thousand documents a month. TypeSafe's own comparison table puts current LLM input prices at $0.20 to $10 per million, with output tokens about five times dearer than input. The same hundred thousand documents on the cheapest of those rates cost $40 before any output tokens; on the dearest, $2,000 before output.
That is the whole commercial case. Not intelligence, not speed. A classify-route-score call that costs cents instead of dollars, and a bill that does not depend on how many tokens the model chooses to say back.
One Hacker News commenter, jrickert, guessed that 40 to 70 per cent of the LLM calls in a typical pipeline are this shape. I would not put a number on it without looking at your bill, but it is the right question, and it is the one to answer first.
The claims, labelled
| Claim | Vendor's figure | Status |
|---|---|---|
| Input price | $0.042 per million tokens | Published; vendor says it "can't prove it isn't subsidized" |
| Output price | Free, "too cheap to meter" | Published |
| Latency | 70–500 ms end to end, against 3–329 s for frontier models | Vendor-measured, "from our laptops on the West Coast" |
| Workflow evals | 193.6x faster, 444.6x cheaper | Vendor's own workflows, written by its own team; vendor calls these "the higher end of real world gains" |
| Type errors | 0% | By construction, "not empirical"; the schema is fixed before the call |
| Public benchmarks | None | Not published, by the vendor's own choice |
| Enterprise terms, SLA, data residency | None | Not published; early access by waitlist |
The latency line drew the sharpest reply on the thread, and it is a fair one:
"70-500ms vs 3-329 seconds" are apples-to-oranges unless the LLM baseline is doing comparable work. If Jev is skipping generation entirely for a narrow structured task, of course it's faster.
— ramon156, on Hacker News
Correct, and it does not matter for the buyer. Nobody is proposing Jev for the work that needs a paragraph back. The comparison that matters is the same narrow task on both models, with the LLM forced into structured output, which is what TypeSafe says its workflow evals do. Those evals are the vendor's. Run your own.
Monday, then. Four things, in order.
Join the waitlist; early access opened on the 15th and there is no other route. Pull last month's model invoice and tag every call that is a classify, route, score, extract or validate. That is the addressable spend, and it is the figure to carry into any conversation with TypeSafe or with a client who is paying you per document. Build the acceptance test before the key arrives: two hundred documents your team has already labelled, run through the schema, and the check is calibration, not accuracy alone. If the model says 90 per cent confidence, it should be right nine times in ten on that slice; if it is not, the confidence score is decoration and the whole pitch goes with it. Monday's post on this site made the same argument for agents: test before you sign.
And sign nothing annual. There are no enterprise terms, no SLA and no deprecation policy on the page, and one commenter's complaint about frontier vendors, that "deprecation cycles" make large-scale automation expensive to maintain, applies twice over to a two-year-old company with one model. A Pakistani house selling document processing to a German client will also be asked where the data goes; TypeSafe says its service is on the US West Coast and says nothing else. Until that line exists, Jev is a cost line for your own tooling, not something to put in a client's architecture diagram.
My call, for what it is worth. The price is the news, and it exposes a gap of five to two hundred times in the one bucket of API spend where volume actually sits. I expect TypeSafe to hold $0.042 through the first quarter of 2027, because it says it expects the price to fall, not rise, and a launch price that moves up is a launch that failed. I also expect at least one of the two large API vendors to publish a structured-decision or classification tier priced under $0.10 per million input tokens before the end of December. If neither happens, the market disagrees with me about how much of your bill is a classifier.
Sources
- Introducing System One Models & Jev, Diogo Almeida, TypeSafe AI, 15 September 2026 (price, latency, eval and hallucination claims, and every quoted caveat)
- Hacker News discussion, 15–16 September 2026 (comments by ramon156, jrickert, bfeynman and others)
- Jev: The Model That Killed Chat GPT's Core Idea? RLCD Explained, Fahd Mirza (YouTube; image)
- Is Classification the Future of AI? and Meet Jev: The AI Built to Make Decisions, vogel (YouTube; images)
- Jev AI Is INSANE… 193× Faster Than LLMs?!, AI WITH Rithesh (YouTube; image)
- Earlier on this site: AI agent acceptance test: the free chess honeypot (14 September)