Jev turns a catalogue ID from a prompt constraint into part of the decision domain

Independent developer Flavio Copes, reiterating TypeSafe's stated performance levels, wrote that most calls to Jev complete in about 100 milliseconds.

TypeSafe emerged after two years in stealth, backed by $40 million in seed funding led by DCVC, to launch its first model, Jev.

Jev is a text-only model that machines can use natively to make decisions inside software; developers send structured questions and get typed decisions with calibrated probabilities.

Jev charges only for input, at $0.042 per million input tokens, with output free.

OpenAI's GPT-5 Nano is priced at $0.05 per million input tokens.

TypeSafe rejects public benchmarks and optimizes for intelligence per dollar, and Diogo Almeida has publicly rejected the idea of a generic JevBench benchmark.

Almeida is ex-OpenAI where, according to TypeSafe, he co-invented RLHF and InstructGPT, the methods behind ChatGPT and GPT-4.

Jev's core innovation is Reinforcement Learning for Calibrated Decisions, a novel unpublished technique that optimizes for answers with epistemically honest probabilities on System One tasks.

In the generative approach selecting a valid identifier is a prompt requirement; with a bounded Choice it is part of the decision domain, which the write-up calls a much stronger contract.

Flavio Copes wrote: "You add it where code needs one judgment, while the rest of the product stays ordinary code."

Jev's input price is about 16 percent below GPT-5 Nano's per input token.

Simon Willison wrote that Jev represents a regression further towards black box machine learning systems: you can put in all the text you want, and the only thing you get back is a floating point number.

Willison wrote: "I really hope nobody uses Jev to rank job applicants."

In Willison's experiment scoring every city in the San Francisco Bay Area on a yes/no "Good city?" question, Jev rated Cupertino top and East Palo Alto bottom.

Willison argues evals and structured experiments are even more important for Jev than for regular LLM projects, and that Jev is cheap enough that running hundreds or thousands of experimental prompts costs just a few cents.

Vercel said that within 24 hours of Jev launching on its AI Gateway, nearly 13% of paid teams were using it, twice the share of the GPT-5.6 family and more than six times Fable 5.1's share.

Almeida posted on X on September 15: "Look at how LLMs generate sequentially, which is great for a natural conversation, but totally useless for computers."

Before Jev, the Invoice Assistant text-analysis stage sent one Gemma 12B call that extracted supplier, invoice number and amount, classified supplier evidence, selected the property, and explained the recommendation with a note.

Part of the original Invoice Assistant prompt said: "Select exactly one property ID present in the supplied catalogue. Do not invent, rewrite or normalize property IDs."

Another part of the prompt said to classify supplier evidence with exactly one qualitative supplierEvidenceStatus: EXPLICIT_TEXT, STRONG_TEXTUAL_IDENTITY, INSUFFICIENT.

Of the pre-Jev design, the write-up says: "It was not disastrously slow. It was not unreliable enough to force a redesign."

The Invoice Assistant pipeline routes to a vision model when supplier evidence from text is insufficient.

Willison has been experimenting with Jev for search reranking, fetching 100 likely matches with BM25 and having Jev score those candidates for relevance against the original query.

Almeida's launch post lists headline stats for Jev as 20-200x faster, 40-400x cheaper with output tokens free, and frontier composable intelligence optimized for decisions.

Almeida claimed System One models output decisions with probabilities and confidence instead of words and "can't hallucinate" because they are "a lot more like code".

For a 20x speedup against a 100ms Jev call, the replaced call must take about two seconds; for 200x, about twenty seconds.