Jev is an AI model released by San Francisco startup TypeSafe AI on 15 September 2026 that returns typed, probability-calibrated decisions instead of generated text (TypeSafe, Introducing System One Models & Jev, 15 September 2026). Within a day of arriving on Vercel’s AI Gateway, nearly 13% of Vercel’s paid teams had tried it — more than twice as many paid teams in its first 24 hours as any previous model launch on the service (Vercel, via Fortune, 8 October 2026). OpenAI shipped a competing Decisions API on 6 October 2026, and TypeSafe announced an $870 million Series A at a $7.5 billion valuation on 9 October 2026 (TypeSafe, TypeSafe A raises Series AI, 9 October 2026).
What is Jev?
Jev is a transformer-based model built for decisions inside software: a developer sends text describing a decision — classify this message, score this candidate, pick one of these options — and Jev returns a typed value with a calibrated confidence estimate that code can branch on. TypeSafe calls this model class “System One Models,” positioned as machine-native intelligence in contrast to chat-optimized LLMs (TypeSafe launch post, 15 September 2026). TypeSafe founder Diogo Almeida describes Jev as “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out” (TypeSafe launch post).
Jev was trained with a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for calibrated probabilities on decision tasks, in contrast to RLHF’s optimization for human-preferred text (TypeSafe launch post). Because Jev does not generate strings, output structure is schema-guaranteed: TypeSafe reports a 0% structured-output and tool-call error rate as a guarantee of its output schema, not an observed benchmark rate (TypeSafe homepage comparison, accessed 23 October 2026).
How does Jev differ from an LLM?
An LLM samples one token at a time, each conditioned on the last, and software must parse the completed string before acting on it. Jev emits every output in a single parallel query against a developer-defined schema, which is where TypeSafe claims much of its speed advantage originates (TypeSafe launch post, 15 September 2026). TypeSafe’s published comparison:
| Dimension | Jev (TypeSafe) | Frontier LLMs (GPT-6, Claude, Gemini) |
|---|---|---|
| Output | Typed decisions with calibrated confidence | Generated text, token by token |
| Sampling | Parallel — all outputs in one query | Sequential — one token at a time |
| Input price | $0.042 / 1M tokens; output free | $0.20–$10 / 1M input; output ≈5x input |
| End-to-end latency | 70–500 ms | 3–329 s for frontier models |
| Type errors | 0% — schema-guaranteed | 0.49%–12.6% structured-output and 0.34%–22.1% tool-call error rates |
| Best suited for | Classification, routing, scoring, verification | Chat, writing, code generation, open-ended reasoning |
Latency, price, and error figures are TypeSafe’s published claims as of September–October 2026. The LLM latency range cites the public llm-benchmarks.diegoromero.es tracker; error rates come from an OpenRouter historical snapshot dated 10 September 2026, which TypeSafe notes may route complex queries toward stronger models (TypeSafe launch post and homepage).
How much does Jev cost?
Jev is priced at $0.042 per million input tokens ($42 per billion), with output tokens free (TypeSafe homepage, accessed 23 October 2026). TypeSafe’s published price comparison, September 2026:
| Model | Input $/1M tokens | Output $/1M tokens |
|---|---|---|
| GPT-6 Astra (OpenAI) | $10.00 | $50.00 |
| Claude Fable 5.1 (Anthropic) | $10.00 | $50.00 |
| Claude Opus 5 (Anthropic) | $5.00 | $25.00 |
| GPT-5.6 Sol (OpenAI) | $4.00 | $20.00 |
| GPT-5.6 Terra (OpenAI) | $2.00 | $12.00 |
| Claude Sonnet 5 (Anthropic) | $2.00 | $10.00 |
| Claude Haiku 4.5 (Anthropic) | $1.00 | $5.00 |
| GPT-5.6 Luna (OpenAI) | $0.20 | $1.20 |
| Jev (TypeSafe) | $0.042 | Free |
Paolo Rosson, head of applied AI at bookkeeping software company Dext, measured Jev at roughly seven cents per 1,000 classified changes in a project screening proposed software changes (Fortune, 8 October 2026).
How fast is Jev?
End-to-end response time is 70–500 ms per call, against 3–329 seconds for frontier LLMs on decision-shaped workloads (TypeSafe launch post, citing llm-benchmarks.diegoromero.es). On TypeSafe’s published workflow evaluations — scored against reference answers averaged from GPT-6 Astra and Claude Fable 5.1 — Jev completed System One-shaped workflows 193.6x faster and 444.6x cheaper than LLM baselines (TypeSafe, evals.typesafe.ai, accessed 23 October 2026). TypeSafe states these gains sit “on the higher end of real world gains” and that the workflows were authored by TypeSafe staff, though their content was not drawn from Jev’s training distribution (launch post).
A Jev-driven Doom bot published by TypeSafe issues about 10 queries per second at roughly $7 per hour of play (TypeSafe launch post, 15 September 2026).
Why is the model called Jev?
Jev is named for economist William Stanley Jevons, who argued that more efficient steam engines would increase total coal consumption by making coal-powered work cheaper — the Jevons paradox. TypeSafe’s stated thesis is that every order-of-magnitude drop in the cost of intelligence unlocks orders of magnitude more automated use cases (TypeSafe launch post FAQ, 15 September 2026). “System One Models” is drawn from psychologist Daniel Kahneman’s Thinking, Fast and Slow: fast, intuitive System 1 judgment, in contrast to slow, deliberate System 2 reasoning (TypeSafe launch post FAQ).
Who built TypeSafe AI?
TypeSafe AI was founded in San Francisco in 2024 by Diogo Almeida, Sasha Sheng, and Erik Gafni (company announcement). Almeida joined OpenAI in 2020 and worked on the RLHF methods behind ChatGPT; OpenAI CEO Sam Altman told him to stop his RLHF work and pursue the decision-model idea instead (Almeida, LinkedIn post, via Fortune, 8 October 2026). Sasha Sheng came from a research engineering role at Meta (company announcement). TypeSafe spent two years in stealth before Jev’s release, and Almeida initially expected reliability work to take a week: “Making the models reliable was so much harder than expected” (Fortune, 8 October 2026).
How much has TypeSafe AI raised?
TypeSafe AI has disclosed two rounds totaling $910 million: a $40 million seed led by DCVC (Fortune, 8 October 2026), and an $870 million Series A at a $7.5 billion valuation led by Andreessen Horowitz with participation from Sequoia Capital, existing investor DCVC, and angel investors, announced 9 October 2026, with a16z general partner Martin Casado joining the board (TypeSafe, TypeSafe A raises Series AI, 9 October 2026; Bloomberg and TechCrunch, 9 October 2026).
| Date | Event | Source |
|---|---|---|
| 2024 | TypeSafe AI founded in San Francisco | Company announcement |
| 15 Sep 2026 | Jev released in early access after two years in stealth | TypeSafe launch post |
| Sep 2026 | Jev arrives on Vercel’s AI Gateway; nearly 13% of Vercel’s paid teams try it within a day — more than twice any prior model launch | Vercel via Fortune |
| 6 Oct 2026 | OpenAI launches Decisions API in public beta | OpenAI docs; Fortune |
| 9 Oct 2026 | $870M Series A at $7.5B valuation; Andreessen Horowitz leads | TypeSafe; Bloomberg; TechCrunch |
What did OpenAI ship in response?
OpenAI launched a Decisions API on 6 October 2026 — three weeks after Jev’s release — as a dedicated POST /v1/decisions endpoint in public beta, running its gpt-6-luna model and returning typed answers about 10x faster than the same model through the general-purpose Responses API (OpenAI, Decisions API guide, accessed 23 October 2026). The endpoint supports three question types: predicate (a probability from 0 to 1 that a condition is true), choice (one of a supplied set), and score (a probability-weighted rating against ordered levels) (OpenAI docs). Decisions accepts images as well as text; Jev accepts text only (OpenAI docs; Fortune, 8 October 2026).
OpenAI lists Decisions input at $0.10 per million tokens, and neither Decisions nor Jev charges for output tokens (Fortune, 8 October 2026). OpenAI’s Nikunj Handa credited Jev with “inspiring this whole thing” and said the Decisions API was not on OpenAI’s development roadmap four weeks before launch (Handa, Latent Space podcast interview, via Fortune, 8 October 2026).
What do developers report?
Fortune’s 8 October 2026 reporting quotes early Jev users. Aaron Roy, developer of the ingredient-scanning app GlutenOrNot, found that Jev missed important emails in a triage test until he rewrote the questions to ask specifically whether a message concerned his money, children, or medical issues, and now runs Jev for initial classification with frontier LLMs as fallback when Jev is uncertain (Fortune, 8 October 2026). Niels Mouthaan, developer of the Daily time-tracking app, used Jev to flag spelling errors as users typed, a case where near-instant feedback is required (Fortune, 8 October 2026). Ashutosh Mathore, founder of environmental-compliance startup Atlensa, said Jev’s speed made him reconsider “where we may be using LLMs out of habit even when the task is not really a language problem” (Fortune, 8 October 2026).
What can Jev not do?
Jev does not generate text — no chat responses, no code, no written reasoning (TypeSafe launch post, 15 September 2026). It accepts text input only as of 23 October 2026, while OpenAI’s Decisions API also accepts images (OpenAI docs). Choice cardinality tops out at 255 options per query, with a two-stage score-then-choose pattern beyond that (TypeSafe launch post). Accuracy claims are workflow-specific: TypeSafe’s published workflow chart places Jev at roughly 68% accuracy — the lowest cost among compared models, though not the highest accuracy (TypeSafe homepage, accessed 23 October 2026). TypeSafe itself flags that its pricing durability is unproven: “We can’t prove it isn’t subsidized” (launch post).
TypeSafe’s Series A announcement claims “a third of the Fortune 500 are getting their Jev on” (TypeSafe, TypeSafe A raises Series AI, 9 October 2026). The company has published no customer list or usage dashboard supporting that claim as of 23 October 2026.
Sources
- TypeSafe AI — Introducing System One Models & Jev — launch post, 15 September 2026, the primary record for Jev’s architecture, RLCD, pricing, and workflow claims
- TypeSafe AI — TypeSafe A raises Series AI — $870M Series A at $7.5B valuation, 9 October 2026
- TypeSafe AI homepage — live pricing, error-rate, and workflow-accuracy comparisons, accessed 23 October 2026
- TypeSafe workflow evaluations — the 193.6x / 444.6x workflow eval details
- OpenAI — Decisions API guide — endpoint, question types, and
gpt-6-lunaavailability, accessed 23 October 2026 - Fortune — Wen Shao, “Jev, an AI for making quick decisions, has been a viral hit in Silicon Valley. But OpenAI is hot on its heels” — 8 October 2026, Vercel adoption data, developer interviews, and OpenAI response
- Bloomberg — “Andreessen Horowitz Backs Jev Maker at $7.5 Billion Value” — 9 October 2026
- TechCrunch — “The maker of non-text AI model Jev valued at $7.5B just weeks after launch” — 9 October 2026
This reference is updated as Jev, TypeSafe AI, and competing decision APIs ship. The canonical URL does not change. Corrections and additions: editorial@hypogray.com.