Jev: The AI Model That Refuses to Write a Single Word (And Why Developers Are Obsessed)
Every AI launch of 2026 chased better conversation. Then on 15 September 2026, ex-OpenAI researcher Diogo Almeida's startup TypeSafe AI launched Jev — a frontier model that will never write you a sentence. Developers crashed its API within days. Here is how it works, where it fits, and how to try it.
What Jev actually is
Jev is the first public "System One" model — named after Kahneman's fast, intuitive thinking (and economist William Jevons). Give it a block of program state plus typed questions, and it returns structured decisions with calibrated confidence scores in a single parallel pass. No chat, no prose, no parsing. Because outputs are defined in advance, it cannot hallucinate — there is simply no free text to go wrong.
Key numbers (vendor-reported)
- Speed: 70–500ms per decision — 40–200x faster than frontier LLMs on comparable tasks
- Cost: $0.042 per million input tokens, output tokens currently free
- Funding: $40M seed round, ex-OpenAI leadership
How it works: state + 3 question types
A Jev request = one block of state (text, JSON, ticket, log burst) + one or more typed questions, all evaluated in parallel:
| Primitive | You ask | You get |
|---|---|---|
| Choice | Which category? | Selected option + per-option probabilities + confidence |
| Score | How urgent (low/med/high)? | Score + distribution + confidence |
| Noul | Is this statement true? | Probability 0–1, e.g. angry-customer: 0.9 |
Example: send a support ticket as state with category?, urgency? and needs-human? — one call returns billing (0.87), urgency 78, escalate: 0.12. Your code acts on numbers above a threshold and routes the rest to a human. Nothing to parse, ever.
Jev vs LLM: honest comparison
| Jev | Frontier LLM | |
|---|---|---|
| Output | Typed values + probabilities | Free text |
| Latency | 70–500ms | Seconds to minutes |
| Cost profile | ~$0.0004 per decision | $0.03–0.17 per case |
| Hallucination | Impossible by design | Needs guardrails |
| Chat / code / reasoning | Cannot do it | Core strength |
| Explanations | None — a number, not a rationale | Full reasoning chains |
The real pattern emerging: LLM for thinking, Jev for deciding. An agent reasons with Claude/GPT, then calls Jev dozens of times for routing, validation and guardrail checks at negligible cost.
Where you can use it (use cases)
Inside software (strongest fit)
- Ticket triage: classify + score urgency for millions of support tickets
- Agent guardrails: screen every LLM tool-call for jailbreaks or risky actions before it executes
- Model routing: decide per-request whether the job needs GPT-6-class muscle or a cheap mini model
- Lead scoring & fraud flags: calibrated probabilities plug straight into thresholds
Real-time loops
- Game bots (launch demo played Doom at ~$7/hour), drone navigation, simulations reacting 10x per second
- Review scoring at scale — ~50M reviews for roughly $20 at advertised rates
Where NOT to use it
Chatbots, code generation, content writing, legal/medical explanations — anything needing words or reasoning stays with LLMs.
Want Jev ideas for YOUR industry? Tell us your business — school, clinic, shop, factory, services — and we'll map exactly where Jev fits in your workflow, free. Popular requests become our next blog guides.
How to try Jev in 4 steps
- Get access: join early access at TypeSafe AI — waitlist is clearing fast, starting credits around $5 (~120M tokens).
- Pick ONE decision: e.g. "is this enquiry sales-ready?" with 3–5 fixed categories.
- Call the API with state + questions; log the confidence scores for a week without acting on them.
- Set thresholds: auto-act above 0.95, escalate 0.5–0.95 to a human, review the rest. Measure for 30 days, then expand.
Developers can also use it via LangChain's TypeSafeClassifier and gateways like Vercel and Cloudflare, which picked it up within days of launch.
Honest caveat: every benchmark so far is TypeSafe's own — no large independent reproduction exists yet, and the company concedes the $0.042 pricing may be subsidised. Promising, not proven. Watch it with a sceptic's hat on.
The bottom line
Jev splits AI in two: language models for language, decision models for decisions. If your software makes the same small judgements a million times a day, a half-second calibrated number beats a five-second eloquent paragraph — at 1/76th the cost. That boring arithmetic is exactly why developers are obsessed.
FAQs
What is the Jev AI model?
A decision-making model by TypeSafe AI (Sep 2026) that returns typed answers with confidence scores instead of text.
Can Jev replace ChatGPT?
No — it cannot chat, code or explain. It handles fast structured decisions; LLMs handle reasoning.
How much does Jev cost?
$0.042 per million input tokens, output free (vendor-reported) — tens to hundreds of times cheaper than frontier LLMs per decision.
Keep reading:
Building AI agents or automation? Idea Creatorz designs agent stacks — LLM for thinking, decision models like Jev for the fast checks. Let's scope yours.