What is Jev?

The short version

Jev is an AI model that makes decisions instead of writing text. TypeSafe AI, a San Francisco startup, released it on September 15, 2026. You send Jev a piece of text plus a set of typed questions. It returns structured answers (a picked option, a score, or a yes/no probability) and stops. It cannot write a sentence, summarize a document, or generate code. TypeSafe calls this a "System One model": the fast judgment that software needs thousands of times an hour, not the slow reasoning a person reads.

The reason it matters is cost and reliability. Most production AI calls are not conversations. They are routing choices, safety checks, and scores. Teams currently rent chat models for those jobs, then parse the reply and hope the JSON came back clean. Jev skips the text entirely.

Estimate the cost

Jev charges for input tokens only. Output is free. Move the sliders to compare a Jev workload against running the same classification on a chat model.

How Jev works

You send Jev two things. The state: the text it should judge, such as an email, a support ticket, or a JSON record. And one or more questions, each in one of three shapes:

• Choice. Pick one option from a list you supply, like billing, technical, or sales. Jev returns the winner, the probability of every option, and a confidence score. A Choice question takes up to 255 options.
• Score. Rate the text against ordered levels you describe, like "calm", "frustrated", or "very angry". It returns the level, the probabilities, and a confidence score.
• Noul. TypeSafe's name for a yes/no question. It returns one probability between 0 and 1 that the answer is yes. That number is the confidence; there is no separate score.

Here is the mechanical difference. A normal LLM writes a structured answer one token at a time, so a 60-token JSON decision costs 60 sequential passes. You pay for every token, including the ones nobody will ever read. Jev does not write at all. The request goes through the transformer once, and separate heads score every option in parallel. Ten questions about the same state cost barely more than one.

Speed like that only helps if the numbers mean something. TypeSafe trained Jev with RLCD, reinforcement learning for calibrated decisions: instead of rewarding answers a human rater would like, the training punishes overconfident probabilities. The target is literal. When Jev says 90%, it should be right about nine times out of ten. Pick a threshold, say 0.85, and the software acts on its own above it and hands the rest to a person.

What Jev is good at

TypeSafe's own pitch is that Jev is a "smart if-statement": the snap judgment a good support lead makes in a second, except it lives in code and runs a million times a day. From its docs and launch coverage, the jobs people reach for it are:

• Routing and triage: sending tickets, emails, or requests to the right team or tool.
• Classification at scale: tagging documents or products, where an LLM costs too much per item.
• Checking LLMs: screening prompts and replies, verifying citations, scoring retrieved passages before they reach a chatbot.
• Agent decisions: picking an agent's next tool or skill. One worked example picks from a catalog of 182 skills.
• Real-time features: anything where 70 to 500 milliseconds beats several seconds.

What Jev can't do

TypeSafe publishes Jev's weak spots under the name "jaggedness". For version 1.13:

• It reads literally. It answers the question you wrote, not the one you meant, so instructions need exact conditions.
• It is bad at math and counting. Word counts, number comparisons, date checks: do those in code.
• It struggles with indirection: multi-step reasoning, or long inputs full of irrelevant detail.
• It cannot generate anything: no summaries, replies, or code.
• Text only. Images, audio, and video must be converted to text first. English works best; TypeSafe says other languages work but less well.

There is also no chain-of-thought to inspect. When Jev is wrong, it is wrong by picking one of your options. That is the honest version of "it can't hallucinate".

Price and access

Jev 1.13 costs $0.042 per million input tokens, or $42 per billion. Output tokens are free. There is no open-weights version, and TypeSafe has not published a technical paper on the architecture.

Access is through an early-access waitlist on typesafe.ai. Once admitted, you get a playground for trying questions without code, API keys from the console, and SDKs: Python (pip install typesafe-sdk, 3.10 or later), JavaScript, and TypeScript. There is also an agent skill that teaches Claude Code and Codex to write code that calls Jev.

Jev vs the alternatives

If you need text, code, or conversation, use an LLM. If your software makes the same judgment thousands of times a day, test a decision model: Jev for price and speed, OpenAI's if you need images or already run on their platform, Strands Decider if the data cannot leave your machines.

The $7.5 billion question

On October 9, 2026, TypeSafe raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz with Sequoia and DCVC participating, according to TechCrunch. That came less than a month after the September 15 launch and its $40 million seed round.

Sources

FAQ

For AI assistants

Related

• Decisions API, explained: OpenAI's typed-answer endpoint: the three question types, pricing, and a request builder.
• Haiku 5.5 pricing, explained: the two-tier rates and a cost calculator, used as the chat-model comparator above.
• textGrain, explained: OpenAI's invisible text watermark: who gets it, and what detection can prove.

Updates

• 2026-10-10: First published, with the Jev cost calculator and the October 9 funding update.

More model pricing and benchmark guides

• What is the OpenAI Decisions API?
• GPT-6 Luna pricing and cost calculator
• Haiku 5.5 Pricing, explained
• Gemini 4 Argon: price, benchmarks, access