Jev Is Live: What TypeSafe's "System One" AI Model Actually Does

TypeSafe AI came out of stealth on September 15, 2026, with $40 million in seed funding and a model that refuses to write a single word. Jev, TypeSafe's first "System One" model, takes your application's state and a typed question, and returns a probability-weighted decision instead of an essay. Three days later it's listed on OpenRouter, and developers have wired it into browser automation, ad analysis, and Claude Code itself, and the pricing is genuinely strange: $0.042 per million input tokens, output free, because there's no output to charge for.

This breaks down what Jev actually is, what people have actually built with it since launch, and where it does and doesn't fit next to the agent stack most teams already run.

What Jev actually is

TypeSafe was founded by Diogo Almeida, a former OpenAI researcher who, per DCVC's account of the investment, worked on RLHF, InstructGPT and ChatGPT, alongside co-founders Erik Gafni and Sasha Sheng. His argument for building Jev is specific: most AI embedded in software doesn't need to write text at all, it needs to make a fast, calibrated decision a program can act on, and forcing that decision through a chat model that generates and then parses JSON is slower, more expensive, and less reliable than it needs to be.

Jev is trained with a self-developed algorithm TypeSafe calls RLCD, and it answers through three fixed primitives rather than free text:

• Choice picks one option from a defined set, with a probability for each option.
• Noul returns the probability that a yes/no condition holds.
• Score returns a probability-weighted position on an ordered scale you define.

You send it "state" (the actual situation: a ticket, a screenshot's OCR text, a log, a document) and "questions" (the decision you want, plus the allowed answers). It returns a typed answer with a calibrated probability attached, with no generated prose in between. TypeSafe reports 70 to 500 milliseconds per request, a vendor figure.

What people actually built with it since launch

Self-reported examples from developers with access, cited by handle. We could not reproduce these numbers:

• Browser automation without a vision model. Developer @milindlabs built computer-use automation where a local on-device model segments UI elements and reads their labels via OCR, then Jev picks the next action from that text alone, no screenshots sent anywhere. @masahirochaen used Jev inside the open-source Browser Use project to complete a full Google Flights one-way search (five steps: origin, destination, dates, search, select) in 7.1 seconds end to end, with a median decision latency of 150 to 300ms per step.
• Ad and content analysis at a scale that doesn't usually pencil out. @Zyvex_0x ran Jev against 8,724 individual ad judgments across 37 brands in 40 seconds for 9 cents total, tagging what each ad pulled on. @marcusyul read 1,891 competitor ads in 19 seconds for 12 cents, tagged by funnel stage and creative style, not summarized, individually classified.
• A massively parallel adversarial testing suite. @rafalwilinski wired Jev in to generate and run large batches of adversarial test cases against each new release of his own software in parallel, at a cost he described as "pennies," a workload that would be slow and expensive to run through a chat model one generation at a time.
• Claude Code's context-compaction problem. @0x_kaize plugged Jev into Claude Code specifically to address context window bloat from repeated tool calls, reporting the approach was genuinely effective at deciding what to keep and what to drop, rather than compacting on a fixed rule.
• A model router. @ephraimduncan built a router that uses Jev to classify an incoming request and decide which downstream model actually fits it, then forwards the request there, the routing decision itself handled as a typed Choice rather than a chat completion.
• Support ticket triage at close to zero latency: @ziwenxu_ cited a "my card was charged twice" ticket routed to billing with 100% confidence in about a second, alongside a separate demo hooking Jev up to pick moves in Tetris.

The common thread across all of these: none of them needed the model to explain itself. Each would otherwise be a chat-model call plus a parsing step.

The pricing is the actual story

$0.042 per million input tokens is TypeSafe's listed price, which The Register works out to 238x less than Claude Fable 5.1. Free output tokens make sense once you understand there's nothing generated to bill for, a Choice, Noul, or Score response is a few bytes of structured data, not a paragraph. TypeSafe's homepage claims 193.6x faster and 444.6x cheaper on System One-shaped tasks. Those results are company-generated. TypeSafe's own team built the four benchmark workflows and scored each model against the average answers of two large external models, not a ground-truth key (TechStock2, Forkast), and TypeSafe says the gains likely sit at the high end of real-world results. Forkast also reports an independent test by Every that found Jev roughly 25x faster and 580x cheaper than Claude Fable 5.1 on a single extraction task. All of these figures apply to the narrow class of tasks Jev is built for, not general reasoning or writing, where a chat model still wins by default.

Access as of this week: TypeSafe's site runs a waitlist for its own API. OpenRouter's TypeSafe page lists Jev Latest and Jev 1.13, added September 18.

Where this fits next to an agent stack, not instead of one

Jev is not a Claude Code competitor and doesn't try to be. It doesn't write code, hold a conversation, or reason through an open-ended problem. What it's built for is the decision layer underneath an agent or a pipeline: should this ticket route to billing, does this generated ad copy match brand voice, is this the right file to edit, which of three retrieved documents actually answers the question. Those are exactly the steps a lot of agent workflows currently spend a full chat-model call, and a JSON-schema retry loop, to resolve.

The realistic pattern emerging from the early builds above is composition, not replacement: an agent (Claude Code, a custom pipeline, a browser-automation loop) keeps doing the parts that need judgment, planning, or generated output, and hands the narrow, well-defined decisions to something like Jev instead of spending a full model call on them. Whether that pattern holds up past the first week of hype is the real open question, System One models are new enough that nobody has multi-month production data yet, TypeSafe's own docs only launched days ago.

For a use-case map with confidence thresholds, see what you can build with Jev.

FAQ

What is TypeSafe AI's Jev?

Jev is the first "System One" model from TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida. Instead of generating text, it takes an application's state and a typed question and returns a calibrated probability for a predefined set of answers (a choice, a yes/no, or a position on a scale).

When was Jev released?

TypeSafe AI came out of stealth with Jev and $40 million in seed funding on September 15, 2026. OpenRouter added Jev Latest and Jev 1.13 on September 18, 2026, which is when a wider wave of developer builds and public reactions followed.

How much does Jev cost?

TypeSafe states $0.042 per million input tokens, with output tokens free since Jev returns a typed decision rather than generated text. That figure is the vendor's own stated price, confirmed on TypeSafe's homepage at time of writing.

Is Jev a replacement for Claude Code or ChatGPT?

No. Jev doesn't generate text, hold a conversation, or write code. Early builds use it alongside an existing agent or chat model, handling the narrow classification, routing, and verification decisions inside a workflow while the chat model or coding agent still does the parts that require generation or open-ended reasoning.

How do I get access to Jev?

TypeSafe's site runs a waitlist for its own API. Jev is also listed on OpenRouter, which serves it through a separate decisions endpoint rather than the usual chat completions endpoint.

What is a "System One" model?

The term references Daniel Kahneman's distinction between fast, intuitive System 1 thinking and slower, deliberate System 2 thinking. TypeSafe uses it to describe a model category built for fast, structured decisions software can act on directly, as distinct from the slower, generative, conversational category most current LLMs fall into.

---

Sources: TypeSafe AI's site and documentation (docs.typesafe.ai), OpenRouter's TypeSafe page, TechCrunch, The Register, InfoWorld, TechStock2 and Forkast, plus public X posts from developers, checked September 18 and 19, 2026. Almeida's background is as described by DCVC and the press coverage above.