🚀 Jev, the stealth AI model

🤖 New AI model built to make decisions instead of generating text

Author

Veer CH
September 21, 2026

Jev’s Paradox:

The internet is verrry exciiite about a new AI model called Jev that was released earlier this week. It’s a frontier model from TypeSafe AI (a company founded by one of the creators of ChatGPT) that is 200x faster and 400x cheaper than other LLMs. The catch: it can’t generate text.

That’s right, Jev is not an LLM. The best way to describe it is that it is a general-purpose classifier. It’s designed for speed, making decisions in parallel instead of generating text token by token. You give it a question, a set of answers, and it gives you the probability that each answer is correct in about 40ms - without the giant wall of text.

Here are some of the patterns that are starting to emerge:

  • Speculative Fanout: Ask multiple questions in a single query and get all of them answered in parallel. For example: triage a support ticket and assess its type, severity, frustration level etc.

  • Confidence Gated Routing: Jev can generate confidence scores that can be used to route to different actions. If you are highly confident you can take the action directly, or if you have low confidence, you can add a user prompt to clarify.

  • Composite Scoring: Give Jev a rubric, and it can give a composite score to all of your criteria. The canonical example here is automated resume screening.

  • Intent Routing: Intent is ambiguous, but often the decision tree of how to respond is deterministic. This means Jev can sit in front of complex queries, classify them and route them to the appropriate response.

Bottom line: Developers were tired of the text gods doing everything for them and getting an outrageous bill at the end of the month. So it’s no surprise that Jev appears to be the missing AI primitive that developers actually want to work with.

Jev is a frontier AI model from TypeSafe AI that returns typed, probabilistic decisions instead of generated text. You send it program state and typed questions; it answers all of them in a single parallel pass in 70 to 500 milliseconds, at $0.042 per million input tokens with output free. It cannot produce a malformed value, because you define the answer space before you ask. It launched two days ago, out of two years of stealth!

Jev is built for decisions, so I gave it one: which sentence on this page actually answers what I mean?

Introducing Needle, a new way to FIND (⌘F).

Cmd+F searches for words. Needle searches for meaning. Ask "what happens if I cancel?" and Jev scores the page's source sentences, picks the passages that match your intent, and highlights them in the original text.

No generated answer. No AI summary. Straight to the source.

Needle is an open-source Chrome extension with a React playground for searching your own text. Apache-2.0.

Jev API is free on Vercel AI Gateway until September 25. So if Jev FOMO has been all over your feed, you now have a working app to clone and a few days to run the model without paying for the calls.

Vercel says Jev became the fastest-adopted model in AI Gateway history, reaching nearly 13% of paid teams in 24 hours. That’s more than 2x the GPT-5.6 family and over 6x Fable 5.1.

Jev cannot generate text. It makes the tiny decisions around them. Needle is our first swing at putting that speed somewhere you can use today.

🤖 Get started with Jev for free:

Jev, the stealth AI model that went viral and had people scrambling for access, is now open to everyone. No waitlist. Start here.  It also supports several typed questions about the same state in one call. That makes it fit routers, judges, guardrails, and handoff checks where the output space is already known.

Jev is inside Postgres now. Zachi's pg-jev extension lets you write queries like WHERE jev(people, 'the name is European'); the database streams rows to Jev, asks one yes/no question per row, and returns the ones that pass. The public demo caps each read-only query at 20 seconds and 2,500 model-judged rows, so no, it is not a magic shortcut through a million-row production table.
pg-jev launch thread | pg-jev live demo

Jev can shrink a Claude Code session tokens by up to 90%! Why is agent compaction still a summarization prompt? Tamara Tran's fast-jev-compaction keeps user and assistant text verbatim, asks Jev which old tool calls still matter, and drops or truncates the rest. If Jev errors, lacks an API key, or cannot shrink the context enough, the plugin falls back to Claude Code's built-in summary.
fast-jev-compaction repository

This browser clicks before you finish the sentence. Moritz Kremb's jev-voice-browser sends every partial transcript to Jev, so the browser can act before you finish talking. Say "go back" and the decision can land in roughly 300ms; the app is already moving while the last word is leaving your mouth.
Voice browser thread jev-voice-browser repository

Browser Use with Jev finds flights in ~7 seconds. Browser Use shipped jev-ultrafast, an open-source BU agent where Jev chooses the browser operation and target, calling a text model only when it needs to type. It found a Google Flights itinerary in 7.073 seconds using 17 Jev requests, with median Jev latency of 178ms.
Browser Use Jev ultrafast demo| jev-ultrafast repository

Of course someone gave Jev a trading account. Jarrod Watts built jev-trader, a Bun app where Jev watches the Kuru MON-USDC order book on Monad and answers buy or sell every roughly 300ms block. It posts a limit order one tick inside the touch, cancels the previous one, and does it again when the next block arrives.
jev-trader launch thread | jev-trader repository

Cua built a Jev-like model just for filling forms. Jev introduced the System One idea: skip text generation and return fast, typed decisions that software can act on. CUA-S1-FORMS applies that same pattern to computer use. For every field or checkbox, the tiny 2.8MB model chooses what belongs there, whether to click it, or whether to leave it alone. Cua Driver then acts and checks that it worked.
CUA-S1 launch thread | Cua article on System One models

I Built Non-Autoregressive Decision Models with RL a Year Ago. Then a Frontier Lab Called It a "Breakthrough". This creator was building RL decision models 18 months before Jev.  In March 2025, he published SalesRLAgent, a reinforcement-learned model that produced live sales-conversion probabilities instead of asking an LLM for the judgment. Now he has turned that OPEN-SOURCE family for Jev-style routing, triage, guardrails, and typed decisions across 100+ languages.
Laya research page |Laya GitHub repository

Jev-as-a-Judge for agent evals. Agent evals often use another LLM to grade whether a run succeeded, which means paying a large model to answer a small question. LangChain tried Jev instead and got the same pass/fail verdict as its human labels in all 500 evaluations. The entire Jev run cost $0.34 versus $28.17 for Claude, and its scores bounced around far less between repeats.
LangChain Jev-as-a-Judge article

10-step guide for the Jev-pilled. For everyone staring at an agent stack and wondering where to put Jev, Codila wrote the 10-step version. Worker selection, source checks, continue/stop decisions, and handoff routing become typed questions; writing and tool execution stay with the models already doing them.
Jev engineering roadmap thread

Context7 finds a task where Jev gets smoked. Tested against Gemini Flash and DeepSeek across five parsing jobs. Jev tied three, won page classification 85% versus 56%, then collapsed on crawl-root selection at 27% versus 93%. It was still 10 to 170x faster and 3 to 20x cheaper. Great for bounded page-level calls; keep the bigger model around when the task needs a mental map of the whole site.
Context7 Jev test

This tutorial shows how to use Jev for the fast, repetitive decisions and hand the harder reasoning back to Claude Code. You’ll see it used for skill selection, feedback loops, adversarial testing, code smells, and other checks where you want speed and low cost without burning expensive LLM tokens on every step.