Overview
Jev is a large model from TypeSafe AI that takes the opposite direction from conversational models: it does not chat and it does not write prose, it returns typed decisions with calibrated confidence. The company calls this class of models System One Models, meaning models built for machines to call rather than for people to talk to. Training uses the company's own RLCD method, Reinforcement Learning for Calibrated Decisions, aimed at making the probability attached to each decision reliable enough for software to act on.
The company's benchmark figures are that on System One workflow tasks Jev runs up to 193.6x faster and up to 444.6x cheaper than conventional large models, with a sample decision costing 0.000081 dollars and completing in 0.114 seconds against 0.01388 dollars and 8.566 seconds for the comparison. A third-party hands-on test found Jev ranked second on accuracy in a pre-screening task while costing less, and took the top spot on both accuracy and speed in parallel judgment tasks.
It is worth being clear about the shape of the interface: Jev returns judgments, not text, and is meant to sit inside code as a decision primitive. Every decision carries a confidence estimate, and you set thresholds for when the software acts on its own and when it escalates to a human. Access is by application on the official site, and Vercel was the first platform to integrate it.
Key Features
- Typed decisions only: Returns typed decision results with calibrated probabilities, called like a function rather than prompted like a chat model.
- RLCD training: Reinforcement Learning for Calibrated Decisions, a training method designed so each decision ships with a usable confidence estimate.
- Confidence thresholds: Your code decides when high-confidence results execute autonomously and when low-confidence ones go to human review.
- Very low latency and cost: Official workflow tests show up to 193.6x faster and 444.6x cheaper than conventional large models, with a sample decision at 0.114 seconds.
- Zero-hallucination positioning: No free-text generation; output is confined to a predefined decision space, which removes a whole class of failure modes.
- Vercel first integration: Available by application, with Vercel as the launch integration partner for deployment workflows.
Use Cases
- Content moderation and risk pre-screening: batch-judge violations, route low-confidence cases to humans
- Ticket and email triage: classify type and urgency before anything reaches a queue
- RAG retrieval filtering: judge whether candidate passages are relevant, replacing hand-written rules
- High-frequency lightweight decisions: replace expensive general-model calls for classification workloads
Pros
- Judgment tasks cost and run orders of magnitude better than general large models by the company's own workflow benchmarks
- Calibrated confidence lets you draw the automation-versus-review boundary in code
- Typed output slots directly into workflow logic without parsing free text
- Vercel integration makes the on-ramp short for teams already deploying there
Pricing
Access is by application on the official site. Pricing is per input token; the company's sample puts a single decision at 0.000081 dollars, and a third-party test cites 0.042 dollars per million input tokens with output tokens free. Confirm current pricing on the official site.
Summary
Jev is a decision-only large model from TypeSafe AI. It skips dialogue and prose entirely and returns typed decisions with calibrated confidence, trained with the company's RLCD method. By its own workflow benchmarks it runs up to 193.6x faster and 444.6x cheaper than conventional large models. It fits when classification, moderation or triage workloads are burning general-model budget, and the confidence threshold gives you code-level control over where automation ends and human review begins.
Version History
- Jev Router reaches thousands of tools and models through one MCP entry (2026-10-28): Jev Router can reach thousands of tools and models through a single MCP entry, with a routing layer choosing the specific tool per task instead of configuring each connection in the client.
- Raises $870 million at a $7.5 billion valuation (2026-10-09): TypeSafe AI announced roughly $870 million in funding led by a16z at a $7.5 billion valuation. Its decision model Jev closed the round weeks after launch, with the company saying about a third of Fortune 500 companies have adopted it and that the platform processes trillions of tokens a day. Jev is in early access and requires applying on the official site.
- The official website shows TypeSafeAI version 1.1, including components such as Clock Tool 1.1 and Glider 1.1, and mentions (2026-09-28): The official website shows TypeSafeAI version 1.1, including components such as Clock Tool 1.1 and Glider 1.1, and mentions new content such as RLCD (reinforcement learning for calibrated); existing evaluations only focus on Jev and the CEO calling it a System One model, and do not cover version 1.1 or the new components.
- Simon Willison and Latent Space pick up Jev, TypeSafe CEO calls it a System One model (2026-09-22): Jev became a focal point this week: Simon Willison wrote it up, and Latent Space interviewed TypeSafe CEO Diogo Almeida, who positions Jev as a production-facing System One model rather than a conventional chat model. A hands-on test by Tomer Tunguz reached 80 percent classification accuracy on 98 human-verified production email threads against 47 percent for the production model, at roughly 76 to 209 times lower cost, and he replaced about a quarter of the calls in his own agent with this class of decision model. Another developer built a bookmark search tool called Margin on Jev, and the BestBlogs daily covered related practice on two consecutive days.
- TypeSafe AI releases Jev, a large model focused solely on high-frequency decision-making; the author personally tested its cost-effectiveness in classification judgments. (2026-09-18): TypeSafe AI has launched Jev, a large model focused on high-frequency decision-making. It does not do conversation or text generation, only outputs judgments. It is 20 to 200 times faster than traditional large models, costs $0.042 per million tokens, and output tokens are free. In the author's hands-on testing, Jev ranked second in accuracy on pre-screening tasks and was cheaper, achieving the highest accuracy and fastest speed on parallel judgment tasks. The model uses the RLCD training method to optimize decision calibration. Applications for access can be submitted on the official website at https://typesafe.ai/, and Vercel has already integrated it first.