Image: GitHub / Browser Use
Why it mattersA team on browser-use 0.x has to decide if it wants the new engine. Cloud browsers are the default. The step cap is lower.
A web agent that picks one DOM action per step hits a ceiling fast. Browser Use's 1.0.0rc2, published on 9 November, replaces that loop. The agent now writes Python in one persistent REPL. It calls browser helpers and raw Chrome DevTools Protocol. It runs on Pydantic AI.
The release notes say the library went from 68,815 to 13,154 lines of Python. The public 0.x names most people use still work: Agent, Browser, Tools, run() and final_result().
What changes if you upgrade
The defaults have moved. When the caller does not name a model, the default is now bu-fast on the Browser Use gateway. Browser() is a cloud browser. A local Chromium is Browser('chromium'). The step cap on run() is 100, down from 500. One step is one model request. The CLI's bu and browser scripts are gone, replaced by browser-use.
Install needs the prerelease flag. Plain pip install browser-use still gives 0.13.14 until 1.0.0 is tagged. The project tells users to pin browser-use<1 to stay on 0.x. Every removed 0.x import raises an ImportError that names its replacement and links a migration guide.
The benchmark Browser Use ran
Browser Use published numbers from its own test, BU Bench v2. It ran 60 tasks three times each on cloud browsers, on this release. Browser Use says the 1.0 default scored 62.6 percent on the test. It cost $0.071 and took 585 seconds per task. On the same window, Browser Use Pi scored 61.3 percent at $0.087 and 689 seconds. Browser Use frames the difference as 18 percent cheaper and 15 percent faster. It reports no meaningful difference in score. It used 95 percent task-clustered bootstrap intervals. Every arm ran gpt-6-luna on api.openai.com. The judge was gpt-5.6-luna. Browser Use notes that the default bu-fast model on its own gateway was not in this run.
The known limits the release calls out
The project pins every dependency with == and commits the lockfile. Browser Use says this is a supply-chain policy: a user gets the versions CI tested. The cost: a project needing a different pydantic, pydantic-ai-slim, mcp, google-genai or pillow cannot install 1.0 next to it. Viewport emulation does not work on cloud browsers yet. Another viewport needs Browser('chromium'). The PyPI README and the docs site still show the plain install line. That line keeps giving 0.x until 1.0.0 is tagged.
A team already on browser-use has three calls to make this week. Is the Python-REPL engine worth switching to? Does a cloud browser as the default fit how its agents run? Does a 100-step cap cut off long jobs that were quietly running at the old 500? If any of these matter, test on the release candidate now. If not, pin browser-use<1.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
