On September 17, 2026, a browser agent quietly did something no chatbot has ever done: it searched Google Flights for a one-way ticket from Zürich to London and had an answer back in about 7 seconds, for less than half a cent. It didn't think out loud, it didn't write a single sentence of reasoning, and by design it had no free-text output to hallucinate in the first place. That's the party trick behind Jev, the first model from a new San Francisco startup called TypeSafe AI, and behind jev-ultrafast, the open-source browser-automation agent the browser-use team built on top of it two days after Jev's launch.
This isn't a new ChatGPT, Claude, or Gemini competitor, and LumiChats isn't suggesting anyone switch their daily chat assistant for it, because Jev can't hold a conversation at all. It's something narrower, and for one specific kind of automation, apparently much faster: a model built to make typed, structured decisions instead of generating text. Here's what actually happened, what a "System One" model is, and why the benchmark numbers are turning heads even with real caveats attached.
TypeSafe AI came out of stealth on September 15, 2026 with a reported $40 million seed round led by DCVC and Jev, the first model in a category it calls "System One models" — models that return typed, calibrated decisions instead of free text. Two days later, the browser-use team open-sourced jev-ultrafast, a browser agent built on Jev, which hit Hacker News's front page the same day. In one internal benchmark, a Zürich-to-London Google Flights search that took a median 9.45 seconds and 1,092 browser protocol calls with an earlier setup dropped to roughly 7.1 seconds and 101 protocol calls with jev-ultrafast, at a builder-estimated cost of about $0.0039 per run at list price.
What is a "System One" model, actually?
Every mainstream AI chatbot in daily use today — ChatGPT, Claude, Gemini, Grok — is an autoregressive large language model. Ask it a question and it predicts one token at a time, stringing words together until it has produced a paragraph. That's powerful for writing, reasoning out loud, and conversation, but it's slow, and for well-defined yes-or-no or pick-one decisions, arguably overkill.
Jev works differently. According to TypeSafe's own launch post, you feed it a block of unstructured "state" — say, everything visible on a webpage — plus a set of typed questions defined in advance, such as which button to click and what confidence to assign to each option. Jev evaluates all of those questions in a single parallel pass and returns typed answers with calibrated probability scores. There's no text-generation step and, per TypeSafe and several outlets covering the launch, nothing to parse afterward — the output is already structured data software can act on directly. Some coverage summarized this by saying Jev "never writes a word," and functionally that's accurate: text generation simply isn't part of what it does. Sources: TypeSafe AI blog, DataCamp: Jev, TypeSafe's System One Model
The "System One" category name is a deliberate nod to psychologist Daniel Kahneman's distinction between System 1 thinking (fast, intuitive) and System 2 thinking (slow, deliberate) — TypeSafe positions today's chat-oriented LLMs as closer to System-2-style reasoners and Jev as a fast, intuitive decision engine for narrow, well-specified tasks. The model's own name, Jev, reportedly references 19th-century economist William Stanley Jevons, known for the Jevons paradox — the idea that making something cheaper tends to increase how much of it gets used overall — which fits TypeSafe's bet that a much cheaper, faster decision engine will get invoked far more often than an LLM would be for the same job.
How Jev differs from the chatbots you already use
| Aspect | Typical LLM (ChatGPT, Claude, Gemini, Grok) | Jev (System One model) |
|---|---|---|
| Output format | Free-form text, generated token by token | Typed values plus calibrated probabilities, returned in one parallel pass |
| Best suited for | Conversation, writing, coding, explaining reasoning in prose | High-volume, repeated decisions over a shared state, e.g. "which of these page elements to click next" |
| Can it write free text? | Yes — that's the core capability | No — text generation isn't part of what Jev does |
| Should you chat with it? | Yes, that's the point | No — TypeSafe itself says it's the wrong tool for chat |
The 7-second browser benchmark, explained
Jev launched in early access on September 15, 2026. Two days later, on September 17, the team behind the popular open-source project browser-use released jev-ultrafast on GitHub, where it reached Hacker News's front page the same day and has since picked up well over 15,000 stars. Sources: browser-use/jev-ultrafast GitHub, MarkTechPost coverage
The core idea: instead of feeding a browser agent full-page screenshots or raw HTML and asking a general-purpose LLM to reason about what to click, jev-ultrafast turns every page into a numbered table of interactive elements — buttons, text boxes, dropdowns — and offers a fixed menu of actions: CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, BLOCKED. One request to Jev answers both which action to take and which numbered element to target, in the same pass, with no free-text reasoning step in between.
In the benchmark documented in the project's own repository, a Zürich-to-London search on real Google Flights — picking dates, running the search, reading back results — had a median completion time of 9.45 seconds and required 1,092 browser protocol calls with an earlier, non-Jev setup. With jev-ultrafast, median time dropped to roughly 7.1 seconds (about a 25% reduction) and protocol calls fell to 101 (a roughly 91% cut), including page-load waits. A builder estimate, based on Jev's list price of $0.042 per million input tokens with outputs free, put the per-run cost at about $0.0039. It's worth being precise about what this number is and isn't: it's the team's own benchmark, run as three repeats of one task on one browser profile with fixed settings, and the project's own documentation describes it as illustrative rather than a general reliability benchmark. It searched for flights; it did not complete a booking, and no independent lab has reproduced the numbers yet. Sources: jev-ultrafast performance docs, Gotchaa Lab writeup
| Metric | Earlier (non-Jev) benchmark | jev-ultrafast, Sept 17, 2026 | Change |
|---|---|---|---|
| Median task time | 9.45 seconds | ~7.1 seconds | ~25% faster |
| Browser protocol calls | 1,092 | 101 | ~91% fewer calls |
| Cost per run (list-price estimate) | Not reported | ~$0.0039 | n/a |
| Test conditions | — | 3 repeats, 1 task, 1 browser profile | Self-reported, not a general reliability benchmark |
Who built this, and why it's getting attention
TypeSafe AI is led by Diogo Almeida, a former OpenAI and Google Brain researcher who contributed to the InstructGPT paper, was acknowledged among the contributors to ChatGPT, and co-authored the GPT-4 technical report. Almeida has publicly argued that reinforcement learning from human feedback (RLHF) — a technique he helped establish at OpenAI — optimizes models for what humans prefer to read rather than for whether a decision is actually correct, which he frames as the wrong training target for unsupervised automation. Jev is instead trained with what TypeSafe calls RLCD (reinforcement learning for calibrated decisions), aimed at confidence scores that are honest about how likely an answer is to be right.
TypeSafe's own marketing claims Jev running anywhere from roughly 40x to 200x faster than frontier LLMs on comparable decision tasks, with Tom's Hardware summarizing one specific comparison as 193x faster and 445x cheaper — figures that, like the flight-search benchmark, are self-reported by TypeSafe rather than independently audited. Sources: Tom's Hardware, Forbes: Why Everyone Is Talking About Jev
What this means — and what it doesn't
- It's not a chatbot upgrade. Jev can't write an email, debug code, or explain its reasoning in prose — TypeSafe itself describes it as the wrong tool for chat, code generation, or anything needing a written explanation.
- It's aimed at high-volume, repeated decisions: the kind of narrow, well-specified choice that gets made millions of times a day, where the set of possible answers is known ahead of time.
- The speed and cost numbers are real but self-reported. 9.45s→7.1s and 1,092→101 protocol calls come from the project's own repository, run as three repeats of a single task — promising, but not yet independently verified at scale.
- It's genuinely early. Jev launched in early access on September 15, 2026, and jev-ultrafast was five days old by the time it hit Hacker News's front page — this is a brand-new model category, not an established one.
If you're evaluating Jev or jev-ultrafast for a real project, treat the 7-second and 445x figures as TypeSafe's own marketing numbers to validate against your own workload, not as independently confirmed benchmarks — the project's documentation says as much itself.
01Is Jev a replacement for ChatGPT, Claude, or Gemini?
No. Jev doesn't generate free text or hold conversations — TypeSafe and independent coverage both describe it as the wrong tool for chat, coding, or writing. It's built for narrow, typed decisions inside software, not for talking to people.
02What does "System One" actually mean?
It's TypeSafe's name for a new category of models that return fast, structured decisions instead of generated text — a nod to psychologist Daniel Kahneman's "System 1" (fast, intuitive) versus "System 2" (slow, deliberate) framing of human thinking.
03Is the 7-second flight-search benchmark independently verified?
No. It's a self-reported result from the jev-ultrafast project's own repository, run as three repeats of one task on one browser profile. The project's own docs call it illustrative, not a general reliability benchmark, and no independent lab has reproduced it yet.
04How much does Jev cost to use?
TypeSafe's list price is reportedly $0.042 per million input tokens, with outputs free. The often-cited $0.0039 flight-search cost is a builder's own estimate from that list price, not an actual invoice.
05Can I try jev-ultrafast myself?
Yes — it's open source on GitHub under the browser-use organization, released September 17, 2026. Jev's underlying model, however, is TypeSafe's closed, hosted API, not something you download.
If all of this makes you want to see how today's mainstream chat models — ChatGPT, Claude, Gemini, Grok, and others — actually compare on the everyday tasks people use them for, that's exactly what LumiChats is built for: one place to compare models side by side and chat with several of them without juggling separate subscriptions. Jev isn't one of the assistants you'll find there — it's not built for chat — but it's a useful reminder that "AI model" is becoming a much bigger category than the chatbot most people picture.
