TypeSafe AI launches Jev, a System One model that makes decisions inside software

TypeSafe AI, a San Francisco AI lab founded by former OpenAI researcher Diogo Almeida, came out of stealth on September 15, 2026, with its first model, Jev, and a $40 million seed round led by DCVC. Jev isn’t a chatbot. It belongs to a new class of models TypeSafe calls System One Models: instead of writing text, it answers typed questions — pick an option, give a score, say yes or no — and attaches calibrated probabilities to every answer, so other software can act on the result. Jev is available in early access through a waitlist.

TypeSafe AI and Jev at a glance

What is TypeSafe AI?

TypeSafe AI describes itself as a frontier AI lab building “machine-native” intelligence: AI designed to run inside software, in the background, rather than to hold conversations with people. The company was founded in 2024 and spent about two years in stealth before the launch.

Its chief executive, Diogo Almeida, worked on reinforcement learning from human feedback (RLHF), InstructGPT, ChatGPT and GPT-4 at OpenAI, and TypeSafe credits him as a co-inventor of RLHF and InstructGPT. He has also worked at Google Brain. His co-founders are Sasha Sheng, the chief operating officer and a former research engineer at Meta’s FAIR lab, and Erik Gafni, the chief technology officer, who previously founded Ravel, a company applying multimodal AI to DNA sequencing.

The $40 million seed round was led by DCVC, a venture firm that backs deep-tech companies, and Forbes reported that it valued TypeSafe at $200 million. In the funding announcement, DCVC general partner James Hardiman said TypeSafe is tackling one of AI’s biggest remaining problems: turning capable models into technology developers can reliably build into products.

Almeida’s argument is that chat models have been better than people at conversation for years, yet automation hasn’t followed. Training with human feedback made models superb at pleasing people, TypeSafe says, but also overconfident and unreliable, which is why most AI systems still need a human in the loop. In a Q&A published by DCVC, he said the goal is for “intelligence to be a utility, a basic building block of automation.”

What is Jev?

Jev is TypeSafe’s first public System One Model. Think of it as a function call rather than a chat: you send state — a support ticket, a log entry, an invoice, a record from your database — along with a set of typed questions, and get back structured answers with probabilities. Jev doesn’t generate free text at all.

The names are borrowed. “System One” comes from Daniel Kahneman’s book Thinking, Fast and Slow, which contrasts fast, intuitive System 1 thinking with slow, deliberate System 2 reasoning. Jev is named after the economist William Stanley Jevons: TypeSafe expects cheaper machine intelligence to increase demand for it, as more efficient steam engines did for coal.

How System One Models differ from LLMs

The comparison is based on TypeSafe’s launch post.

How Jev works: questions instead of prompts

A Jev request has two parts: the state the decision is about, and one or more questions. Each question has an ID, a type, instructions and, where needed, the allowed answers. There are three question types:

A Choice question accepts up to 255 options, and a request can carry about 32,000 tokens — roughly 150,000 characters of English — shared between the state and the questions. Jev answers all the questions in a request in parallel, so one call can ask many things at once.

In an example from TypeSafe’s documentation, a customer writes that their running shoes arrived in the wrong size. Asked which team should handle it — returns, shipping or billing — Jev answers “returns” with a probability of 1.0.

The confidence score is the central idea. Developers set thresholds in code: act automatically when confidence is high, and send the case to a person when it isn’t. TypeSafe trains Jev with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD), which optimises for probabilities that honestly reflect how likely an answer is to be right, rather than for the answers human raters prefer. In a calibrated model, higher confidence means higher accuracy.

Price and speed

• Price: $0.042 per million input tokens — $42 per billion — and output tokens are free. TypeSafe says Jev’s input price is 238 times lower than that of Anthropic’s Claude Fable 5.1. It also says it can’t yet prove its pricing isn’t subsidised, and expects prices to fall rather than rise.
• Speed: 70 to 500 milliseconds end to end, which TypeSafe says is 40 to 200 times faster than language models of similar intelligence on System One-shaped tasks. In the demo on its home page, Jev answered in 0.114 seconds at a cost of $0.000081, against 8.566 seconds and $0.013880 for OpenAI’s GPT-5.6 Terra.
• Rate limits: 250,000 tokens per second and 1,200 requests per minute, adjusted dynamically while demand is high.

TypeSafe’s headline claim, 193.6 times faster and 444.6 times cheaper, comes from its own workflow evals. It ran four business workflows — security incidents, agent trace observability, invoice processing and customer service — with each model, using the average answers of OpenAI’s GPT-6 Astra and Claude Fable 5.1 as the reference. TypeSafe itself calls those gains the higher end of what to expect in practice. It also notes that its own team wrote the workflows, and that it measured speeds from laptops on the US West Coast, where its service runs.

For the models Jev is measured against, see our ChatGPT and Claude profiles.

What you can build with Jev

TypeSafe pitches Jev for decisions where hand-written rules are too brittle, but calling a large language model is too slow or too expensive:

• Smart if-statements: classify, route, score or extract inside ordinary code — for example, sending support tickets to the right team.
• Big data: run the same questions over millions of records to turn text into features and insights.
• Real-time applications: answers in around 100 milliseconds make AI usable where the user experience depends on speed.
• Checks on other AI: score, judge and guardrail the prompts, reasoning and outputs of LLMs, including spotting jailbreak attempts.

TypeSafe’s documentation includes recipes for re-ranking search results, double-checking citations, screening LLM inputs and outputs, extracting dates and classifying documents through deep category hierarchies. For the launch, the team also showed Jev playing Doom from a text description of the game state — about 10 decisions a second, which TypeSafe says costs roughly $7 an hour — and racing between Wikipedia pages by choosing which link to follow.

Limitations and criticism

• No text. Jev can’t write, summarise or chat. It only answers questions you define in advance.
• “Zero hallucinations” has a narrow meaning. TypeSafe guarantees that every answer fits the schema, so Jev never returns an invalid type. As The Decoder pointed out, that doesn’t stop it choosing the wrong option from the allowed list — the confidence score is what tells you when to double-check. The Register made a similar point: comparing hallucination rates with chat models isn’t like for like, because Jev’s output isn’t natural language.
• Structured output isn’t new on its own. Language models from OpenAI and others can already return fixed categories and JSON. Jev’s case rests on speed, price and calibrated probabilities, not on structure alone.
• Benchmarks built in-house. TypeSafe created the workflow evals and scored them against other AI models’ answers rather than independently verified ones.
• Known weak spots. TypeSafe’s documentation for jev-1.13 says the model reads questions literally and is unreliable at maths, counting and comparing dates. It also struggles with numeric encodings such as hex values, and multi-step or double-negative questions reduce its accuracy. Irrelevant data in the state can distract it, and it doesn’t treat input as potentially hostile. TypeSafe advises keeping arithmetic in code and testing integrations before rolling them out widely.

How to get access

Jev is in early access: developers join the waitlist on TypeSafe’s website, and the company says it is bringing people off the waitlist as quickly as it can. Once admitted, you create an API key in the TypeSafe console, which also has a playground for testing questions.

• Python: pip install typesafe-sdk (Python 3.10 or newer)
• JavaScript and TypeScript: npm install @typesafe-ai/sdk (Node.js 20 or newer)
• HTTP API: one endpoint for System One requests, with the model IDs jev-latest, jev-preview and the pinned version jev-1.13.0
• Coding agents: a TypeSafe skill for Claude Code, Codex and other agent environments

If you tune confidence thresholds, TypeSafe recommends pinning a model version, so behaviour doesn’t change when the jev-latest alias moves to a new release.

What the launch means

Much of this year’s AI news has been about better assistants for people. TypeSafe is betting on the opposite: that a large share of AI’s value will come from small, fast decisions made inside software, where no one reads the output. If Jev’s speed, price and calibration hold up outside TypeSafe’s own tests, it could make AI practical where calling a chat model on every event is too slow or too costly — routing, moderation, fraud checks and data pipelines.

The sensible next step is the one TypeSafe recommends itself: test Jev on your own decisions, side by side with the model you use today. If you’re choosing an assistant to work with directly rather than a model to build into software, start with our guide to choosing an AI assistant.