Hand Jev a support ticket and three typed questions (which queue does this belong in, how urgent is it, is the customer about to churn) and it returns three answers with probabilities attached, in 70 to 500 milliseconds, for $0.042 per million input tokens. No prose comes back, so there is nothing to parse and nothing to hallucinate outside the schema. The whole product fits in that sentence. A fortnight of coverage has made it sound like a great deal more.

What shipped in September

TypeSafe AI, co-founded by Diogo Almeida after four years at OpenAI on RLHF and InstructGPT, released Jev on 15 September with a $40 million seed led by DCVC. It is closed: no weights, no paper. TypeSafe calls the category “System One Models”. Simon Willison prefers “decision models”, and so do I.

Three days later Nandakishor Mukkunnoth’s Convai Innovations published Laya under Apache 2.0: a 421 million parameter ModernBERT encoder with the same three question types, running on your own hardware with no hosted API. Mukkunnoth had published a related idea, probability-scored decisions during sales calls, on arXiv as SalesRLAgent eighteen months earlier. Laya was one of six reproductions Latent Space counted within two days of the launch. What is new is the packaging: a schema set at request time, calibrated probabilities out, and a price roughly two orders of magnitude under a frontier call. One independent write-up calls much of what teams route through generative models “classification wearing a costume, paid for at generative prices”.

Vercel had Jev in its AI Gateway on launch day. Ten days after a seed that valued it at $200 million, TypeSafe was reported to be fielding offers at $10 billion.

Read the benchmarks before the headlines

The launch numbers come from the vendor. TypeSafe’s “193.6x faster, 444.6x cheaper” figures come from workflow evaluations its own blog says “were made by individuals on our model capabilities team, so some bias could exist”. Independent tests followed within days. Most are one engineer’s repository, but the code is public and they tell a more useful story.

Out of the box, Jev is well ahead. sysone-bench scored Jev at 0.91 and Laya at 0.69 across 1,240 decisions. On the independent Decision Index, where 0 is random guessing, Laya’s own model page reports 5.51 against Jev’s 51.67 and calls Laya “a fast base to specialise, not a zero-shot decision engine”. Tuned on your data, the gap closes: in a package-triage benchmark, fine-tuned Laya matched or beat Jev on four of five tasks, at around 20 milliseconds against Jev’s 140. Jev works out of the box and you rent it; Laya needs your labelled examples and you own it.

How you ask matters as much as which model you pick. In a 2,000-email phishing test, Jev asked one question scored 62.6% against Claude Haiku 4.5’s 81.3%. Split into five narrow questions, with weights fitted on half the emails, it reached 95.0% on the other half. One Hacker News commenter reports that reordering the answer options can shift Jev’s probabilities by up to 20 percent; unverified, but cheap to test before you trust a threshold.

Where they earn their keep

A Vercel engineer who swapped OpenAI’s model for Jev on safety command review reported results 5 to 18 times faster; Bryo AI’s CTO found it slightly less accurate but 10 to 20 times cheaper than Gemini for email classification. The pattern to copy is any point where a full LLM call ends with the answer parsed into an enum. Willison lists spam, labelling, prioritisation and reranking; Vercel adds tool selection and flagging low-confidence outputs for a human.

Routing is the big one: most agent pipelines spend a frontier call deciding what to do next and another doing it. Move the first to a decision model and every step gets cheaper, with a confidence score to set an escalation threshold on.

Where to keep them out

VentureBeat documented an Octomind engineer asking Jev whether to block rm -rf ~/.ssh. It returned a block probability of 0.76; after one fabricated field claiming the user had pre-approved the command, 0.48. LangChain now ships middleware that hides tool output from the classifier, and Pydantic’s guidance is that a Jev guard belongs “alongside deterministic checks, not instead of them”.

A manipulated LLM produces prose a reviewer can read; a manipulated decision model produces a slightly different float that nobody reads. Willison asked Jev to score Bay Area cities as “good” and got Cupertino at the top and East Palo Alto at the bottom, with no schema to catch it. Keep these models away from anything irreversible (deleting, paying, deploying, hiring) unless a deterministic check also has to pass.

How to adopt this in a week

Pick one classification you currently make with a generative call. Ask several narrow questions rather than one broad one. Pull 200 labelled examples from production logs; if you cannot, that is the real gap. Run Jev through Vercel’s Gateway and Laya locally in shadow mode beside the existing call, log all three answers with their probabilities, and compare after a week. Then put the decision model in front only above a confidence threshold chosen from your own data, and send the rest down the old path. Keep logging the probabilities: drift in a float is the only early warning you will get.

The next three months

The category commoditises fast: six reproductions in two days puts the price floor at zero. Before the year is out, expect a decision endpoint from at least one frontier lab, decision calls as a checkbox in every gateway and agent SDK, one benchmark suite becoming the standard everyone quotes, and a public incident with a decision model somewhere in the approval chain. TypeSafe’s valuation is a bet that the model is the hard part. Two weeks in, the hard part looks like a labelled dataset for your own task, and you can start on that today.

If you want a second pair of eyes on where a decision model fits in your pipeline, ask us.