Browser Use launches Cloud v4, claiming web tasks at 3 cents per success

Why it matters

Browser Use is turning a widely adopted open-source project into metered cloud infrastructure. V4 could improve web-agent economics, but its 3-cent claim reflects a narrower test than its production cost data.

Gregor Zunic (@gregpr07) launched Browser Use Cloud v4 on August 7th, pitching a code-writing web agent designed to complete difficult browser tasks at a sharply lower cost. In a four-post thread on X, the Browser Use co-founder and CTO said v4 addresses two persistent constraints for web agents: accuracy and cost. Speed remains unresolved. New accounts receive $15 in credits, according to Zunic. (x.com)

The chart accompanying the announcement puts the Luna-powered v4 agent at roughly an 88% success rate and $0.03 per successful task. Browser Use plotted v3 at about 80% and $1.35 per success, making the claimed v4 cost 45 times lower. V2 appeared at roughly 68% and $2.02 per success. Those figures are company measurements, and the launch graphic does not identify its task set, number of runs or judging method.

Browser Use's current evaluation documentation offers a more complete and less dramatic view. On an internal set of 106 difficult tasks run four times in July 2026, v4 completed 76.47%, compared with 67.25% for v3 and 54.25% for v2. Each run used Claude Opus 4.8, and Browser Use excluded tasks blocked by websites' anti-bot systems. The company says it used the same independent judge across results. (docs.browser-use.com)

The production economics in those docs also differ from the 3-cent figure in Zunic's launch chart. Across the previous 30 days of customer usage with Claude Opus 4.7, Browser Use lists a median cost per successful task of $1.64 for v4, $0.87 for v3 and $0.70 for v2. Median completion times were about five minutes, two minutes and one minute, respectively. Browser Use attributes the higher v4 production cost to customers assigning it harder work.

The difference appears to come largely from the underlying model and workload. Browser Use recommends GPT-5.6 Luna for v4, pricing it at $0.24 per 1 million input tokens and $1.44 per 1 million output tokens. Claude Opus 5 is listed at $6 and $30 for the same token categories. Browser sessions and network traffic carry separate charges. Browser Use says Luna approaches Opus-level benchmark performance while delivering its lowest token price and fastest median response time among models running in v4 production. That makes the 3-cent result plausible for a bounded evaluation, while the company's broader production data shows that complex customer tasks still cost considerably more. (docs.browser-use.com)

A code-driven agent for longer jobs

V4 changes the way Browser Use approaches web automation. V3 observes pages and takes actions step by step. V4 can write and execute code while working across multiple pages or websites, allowing it to collect records, compare products and save results as spreadsheets or other files. Browser Use positions v4 for bulk data collection and long instructions, reserving v3 for simpler jobs where latency matters.

Developers can access the release through the v4 SDK and runs API. Each run creates a browser session and a persistent workspace, unless the developer supplies existing identifiers. Sessions can carry a conversation and live browser across follow-up runs, while workspaces preserve uploaded files, generated data and scripts across separate conversations. Browser Use also allows agents to save tested scripts and repair them when a website changes. (docs.browser-use.com)

Zunic built Browser Use with Magnus Muller (@mamagnus00) after the pair met at a hacker house at ETH Zurich. Zunic studied physics before earning a master's degree in data science at ETH; Muller studied data science there and had worked on machine learning and process automation. They started Browser Use as a weekend experiment in 2024 and released its open-source project that November. Browser Use later joined Y Combinator's Winter 2025 batch. (ycombinator.com)

The open-source Browser Use repository has accumulated more than 108,000 GitHub stars. The managed cloud service gives Zunic and Muller a way to convert that developer distribution into paid browser sessions, model usage, proxies and orchestration. V4 is central to that strategy because high-volume automation depends on the cost of successful outcomes, rather than the price of an individual model call. (github.com)

Browser Use raised a $17 million seed round in March 2025 led by Felicis Ventures, with A Capital, Nexus Venture Partners, Y Combinator, Paul Graham, Liquid2, SV Angel and Pioneer Fund among the participants. The round financed Browser Use's push from an open-source agent framework into hosted infrastructure for production workloads. (browser-use.com)

V4 improves the company's published accuracy on hard tasks, but Zunic's admission on speed defines the remaining constraint. Ten-minute runs can work for research, data collection and background automation. Interactive applications require much lower latency. Browser Use's v3 agent remains its fastest option on identical hard tasks, taking about six minutes compared with roughly 10 minutes for v4 in the company's tests. V4's commercial test will be whether customers accept that delay in exchange for fewer failed runs and lower model costs.