What people are doing with Jev, the model that decides instead of writes.

We just ran Jev on our WebMCP benchmark.

The result: basically broke the benchmark.

Jev + Mercury 2.5 (a fast, low-cost LLM) using WebMCP solved 100% of the tasks at roughly 112× lower model cost than GPT-6 Astra using computer use with code execution. Compared to Astra using screenshot-based computer use, the model cost was 245× lower (!).

We also compared Jev operating the browser with and without WebMCP.

We used Browser Use’s open-source Ultrafast, with some improvements to the harness to make it more reliable across the benchmark.

Jev’s browser-control accuracy on its own was not amazing - adding WebMCP nearly doubled the number of solved tasks, from 25/49 to 49/49, while reducing model cost by 18% (more on why below).

The benchmark and methodology are fully open and reproducible.

Full results: https://webmcp.com/benchmark

A few words on how the Jev + WebMCP harness works and why this is exciting:

Jev receives text as input and a set of discrete options it can choose from. With WebMCP, those options are the tools exposed by the website. At each step, Jev sees the task, the available tools and previous results, then picks what to do next.

The limitation is that Jev can’t generate arbitrary text, which you need for tool arguments. For example, it can choose the search_products tool, but it can’t generate the search query itself.

So we split the work: Jev picks the tool and Mercury 2.5 generates the arguments if needed.

This works well because turns out most of the cognitive load in these tasks is around choosing the right action. The argument generation itself is relatively simple, so we can delegate to a small and very fast model. We used Mercury, which outputs 1,000+ tokens/sec and is very cheap.

The result is a pretty simple combination: Jev for tool selection + Mercury for arguments + WebMCP for the interface. It ends up being very reliable, very fast, and very cheap.

A few words about Ultrafast and why do we think it underperforms:

Without WebMCP, Jev chooses from the page’s controls: which button to click, which field to fill, or which option to select.

But choosing a valid button is different from choosing the right next step. The agent still has to navigate menus, understand forms, recover from errors and recognize when the task is actually complete.

Our hypothesis is that WebMCP makes the decision space much simpler. Instead of figuring out a sequence of clicks through a website, Jev chooses explicit actions that directly advance the task.

@typesafeai itself documents weaker accuracy on questions requiring multiple reasoning steps. WebMCP moves much of that complexity into the website’s tools, leaving Jev with clearer decisions and fewer opportunities to go wrong (in a sense WebMCP "compresses" a sequence of clicks into one tool call).

Our modified Ultrafast setup solved 25/49 tasks - that is a result for our particular implementation and benchmark, not a universal limit on Jev or Browser Use. We are open to more harness optimization to get this result to perform better, feel free to directly contribute to the benchmark here: https://github.com/nekuda-ai/WindTunnel

Browser-use ultrafast: https://github.com/browser-use/jev-ultrafast

TL;DR of my new article: WTF is Jev by @typesafeai, and the 9 things people are already building with it. The thesis: 𝗮 𝗰𝗼-𝗰𝗿𝗲𝗮𝘁𝗼𝗿 𝗼𝗳 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝘀𝗽𝗲𝗻𝘁 𝘁𝘄𝗼 𝘆𝗲𝗮𝗿𝘀 𝗶𝗻 𝘀𝘁𝗲𝗮𝗹𝘁𝗵 𝗼𝗻 𝗮 𝗺𝗼𝗱𝗲𝗹 𝘁𝗵𝗮𝘁 𝗰𝗮𝗻𝗻𝗼𝘁 𝘄𝗿𝗶𝘁𝗲 𝗮 𝘀𝗲𝗻𝘁𝗲𝗻𝗰𝗲, 𝗮𝗻𝗱 𝗶𝗻𝘀𝗶𝗱𝗲 𝟳𝟮 𝗵𝗼𝘂𝗿𝘀 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿𝘀 𝘄𝗶𝗿𝗲𝗱 𝗶𝘁 𝗶𝗻𝘁𝗼 𝗲𝘃𝗲𝗿𝘆 𝗰𝗵𝗲𝗮𝗽 𝗷𝘂𝗱𝗴𝗺𝗲𝗻𝘁 𝗰𝗮𝗹𝗹 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗺𝗮𝗸𝗲𝘀.

Think AI multiple choice, not AI essay writing. It doesn't chat. You hand it app state plus a typed question, it hands back a decision with a probability attached. 𝟯𝟭.𝟰𝗠 𝘃𝗶𝗲𝘄𝘀 on the launch post in two days (@CompleteSkeptic, who co-invented RLHF). I ran @slashlast30days on it 11 times, then checked every big post by hand.

🌐 𝗔 𝘁𝗶𝗻𝘆 𝗼𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲 𝗯𝗿𝗼𝘄𝘀𝗲𝗿 𝗮𝗴𝗲𝗻𝘁 𝗳𝗼𝘂𝗻𝗱 𝗳𝗹𝗶𝗴𝗵𝘁𝘀 𝗶𝗻 𝟳 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗳𝗼𝗿 $𝟬.𝟬𝟬𝟯𝟵. New action space every step, DOM as state, Jev picks the click, a small LLM only wakes up to type. The Browser Use founder built it (@gregpr07, 7.2K likes, 1.8M views) and had to note the video is 1x speed

🧹 The sleeper: instant compaction. Score every tool call, drop the junk, skip the summarization prompt entirely. "𝘪𝘯 2026, 𝘸𝘩𝘺 𝘪𝘴 𝘤𝘰𝘮𝘱𝘢𝘤𝘵𝘪𝘰𝘯 𝘴𝘵𝘪𝘭𝘭 𝘢 𝘴𝘶𝘮𝘮𝘢𝘳𝘪𝘻𝘢𝘵𝘪𝘰𝘯 𝘱𝘳𝘰𝘮𝘱𝘵?" asked @tamarajtran, 5K likes, then shipped the answer that afternoon. Run as a Claude plugin it took a session 𝗳𝗿𝗼𝗺 𝟭𝗠 𝘁𝗼𝗸𝗲𝗻𝘀 𝘁𝗼 𝟴𝟲𝗞 𝗶𝗻 𝗼𝗻𝗲 𝘀𝗲𝗰𝗼𝗻𝗱 (@altryne). Diogo's reply: "𝘧𝘳𝘦𝘦 𝘤𝘰𝘥𝘪𝘯𝘨 𝘢𝘨𝘦𝘯𝘵𝘴 𝘧𝘳𝘰𝘮 𝘥𝘦𝘴𝘪𝘨𝘯𝘪𝘯𝘨 𝘢𝘳𝘰𝘶𝘯𝘥 𝘵𝘩𝘦 𝘒𝘝 𝘤𝘢𝘤𝘩𝘦"

🛡️ Vercel put it in production as the safety reviewer in fx auto mode. 𝗨𝗽 𝘁𝗼 𝟭𝟴𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 𝗮𝘁 𝗽𝟵𝟱 𝗮𝗻𝗱 𝗺𝗼𝗿𝗲 𝗮𝗰𝗰𝘂𝗿𝗮𝘁𝗲 than the model it replaced, per @rauchg, 3.7K likes. LangChain open-sourced the same idea the next day as AutoModeMiddleware. The closed danger classifier inside every coding harness is now a 100ms primitive

🚦 Model routing as middleware instead of a paragraph in a system prompt. About a dozen lines, probabilities left in agent state so you can audit the choice. The LangChain writeup by @sydneyrunkle is the cleanest how-to-wire-it piece anyone has published

🔎 RAG precision, solved the dumb way: retrieve as usual, run Jev on every chunk, delete the irrelevant ones. "𝘢𝘭𝘴𝘰 𝘥𝘪𝘥 𝘢𝘯𝘺𝘰𝘯𝘦 𝘳𝘦𝘢𝘭𝘪𝘻𝘦 𝘫𝘦𝘷 𝘴𝘰𝘭𝘷𝘦𝘥 𝘱𝘳𝘦𝘤𝘪𝘴𝘪𝘰𝘯 𝘪𝘯 𝘙𝘈𝘎?" (@kushbhuwalka, 416 likes)

🎮 Minecraft in real time: 𝗝𝗲𝘃 𝗿𝗲𝗮𝗰𝘁𝘀, 𝗚𝗣𝗧-𝟲 𝗔𝘀𝘁𝗿𝗮 𝗽𝗹𝗮𝗻𝘀, and they fight multiple zombies at once (@wuyang_zhou). A launcher that reads intent on every keystroke in about 100ms (@dabit3). TypeSafe's own demo is Doom at 10 decisions a second, roughly $7 an hour

📬 Email triage at scale: 1,500 emails in batches of 100 with 8 workers, 60,996 views on the demo. "𝘞𝘦 𝘰𝘯𝘭𝘺 𝘩𝘢𝘷𝘦 𝘢 𝘣𝘢𝘭𝘢𝘯𝘤𝘦 𝘰𝘧 $5 𝘥𝘰𝘸𝘯 𝘩𝘦𝘳𝘦, 𝘸𝘩𝘪𝘤𝘩 𝘫𝘶𝘴𝘵 𝘴𝘩𝘰𝘸𝘴 𝘩𝘰𝘸 𝘤𝘩𝘦𝘢𝘱 𝘵𝘩𝘪𝘴 𝘮𝘰𝘥𝘦𝘭 𝘪𝘴"

🗂️ 𝟳𝟳𝟳 𝗷𝘂𝗱𝗴𝗺𝗲𝗻𝘁𝘀 𝗶𝗻 𝘂𝗻𝗱𝗲𝗿 𝟬.𝟳 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗳𝗼𝗿 𝗮 𝗾𝘂𝗮𝗿𝘁𝗲𝗿 𝗼𝗳 𝗮 𝗰𝗲𝗻𝘁. Every's head of evals asked 21 questions of 37 documents in one request, and that is what came back

🧪 Jev in your browser: Reflex, a Qwen model doing structured decisions on WebGPU, built at Shopify by @kshetrajna and passed around by @tobi. Three independent clones inside 72 hours. 𝗧𝗵𝗲 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 𝗶𝘀 𝘁𝗵𝗲 𝗶𝗻𝘃𝗲𝗻𝘁𝗶𝗼𝗻, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝘄𝗲𝗶𝗴𝗵𝘁𝘀

🔌 Already behind the gateways you use: @vercel AI Gateway inside 48 hours (2,341 likes, the company's second-biggest post), Cloudflare, and @OpenRouter in beta

💸 𝟱,𝟬𝟬𝟬 𝗿𝗲𝗾𝘂𝗲𝘀𝘁𝘀 𝗳𝗼𝗿 𝗮𝗯𝗼𝘂𝘁 $𝟮. That was one developer counting his bill on day one (@MichaelLee04, 3,060 likes). Input is $0.042 per million tokens. Output is free

🧨 The honest part: Every's second test came out 𝟮𝟱𝘅 𝗳𝗮𝘀𝘁𝗲𝗿, 𝗻𝗼𝘁 𝟮𝟬𝟬𝘅, and Jev caught 6 of 7 planted defects to Fable 5.1's 7. The HN launch thread (1,863 points) spent most of its length on "can't hallucinate." Top critical comment: "𝘪𝘵 𝘤𝘢𝘯'𝘵 𝘦𝘮𝘪𝘵 𝘢𝘯 𝘪𝘯𝘷𝘢𝘭𝘪𝘥 𝘵𝘺𝘱𝘦, 𝘣𝘶𝘵 𝘪𝘵 𝘤𝘢𝘯 𝘴𝘵𝘪𝘭𝘭 𝘦𝘮𝘪𝘵 𝘢 𝘤𝘰𝘮𝘱𝘭𝘦𝘵𝘦𝘭𝘺 𝘸𝘳𝘰𝘯𝘨 𝘷𝘢𝘭𝘪𝘥 𝘷𝘢𝘭𝘶𝘦." Diogo called the "it's a zero-shot classifier" read "𝘷𝘦𝘳𝘺 𝘢𝘤𝘤𝘶𝘳𝘢𝘵𝘦!" And the biggest Reddit thread is someone who open-sourced the same architecture a year ago, 1,568 upvotes. Top reply: "𝘉𝘶𝘵 𝘥𝘪𝘥 𝘺𝘰𝘶 𝘱𝘰𝘴𝘵 𝘪𝘵 𝘴𝘢𝘺𝘪𝘯𝘨 𝘪𝘵'𝘴 𝘵𝘩𝘦 𝘯𝘦𝘹𝘵 𝘣𝘪𝘨 𝘵𝘩𝘪𝘯𝘨? 𝘙𝘰𝘰𝘬𝘪𝘦 𝘮𝘪𝘴𝘵𝘢𝘬𝘦"

Bonus: the name is not Kahneman. It's William Stanley Jevons, of Jevons paradox. Make a resource cheaper and people consume far more of it. Naming your decision model after that is a thesis statement.

𝗞𝗲𝗲𝗽 𝘁𝗵𝗲 𝗯𝗶𝗴 𝗺𝗼𝗱𝗲𝗹 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗮𝗻𝗱 𝘄𝗿𝗶𝘁𝗶𝗻𝗴. 𝗨𝘀𝗲 𝗝𝗲𝘃 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗿𝗮𝗽𝗶𝗱-𝗳𝗶𝗿𝗲 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀 𝗶𝗻 𝗯𝗲𝘁𝘄𝗲𝗲𝗻. That's the whole article.