browser-use 1.0 deletes 80% of its code for a Python REPL
The web agent library was rewritten around a persistent Python REPL on Chrome DevTools Protocol, with cloud-first defaults and a benchmark that the team says overstates its gain.
What shipped on 10 November
On 10 November the browser-use project published browser-use 1.0.0, and the headline number in the browser-use 1.0 release notes is a deletion: 136,866 lines removed, roughly 80% of the library, with the team saying it started over rather than refactored. The agent now writes Python into one persistent REPL that runs straight on Chrome DevTools Protocol. There is no DOM serializer and no action schema between the model and the browser. The post calls the result the state-of-the-art web agent.
The loop described in the release is short. The model writes a cell such as new_tab(url); bu.click(41); js(...), the REPL runs it on Chrome, the output goes back to the model, and finish(result) ends the run. Installation is uv add 'browser-use==1.0.0' plus a BROWSER_USE_API_KEY from cloud.browser-use.com/new-api-key. The example in the post is an Agent('Find the number of stars of the browser-use repo') call whose run_sync().final_result() returns the answer. The team asks users to pin with ==.
Defaults that moved to the cloud
Several 0.x defaults changed, and they matter more than the line count. The llm argument used to be required and now falls back to bu-fast on the Browser Use gateway. Browser() used to launch local Chromium and now starts a Browser Use Cloud browser, with Browser('chromium') for a local one. Tasks on localhost or private IPs need that local browser, and the release says a warning tells you so. Instead of a key per model provider, there is one Browser Use key.
Other changes worth checking before you upgrade: ultrafast mode is the default on bu-fast and bu-ultrafast, and mode='default' opts out. run() now defaults to max_steps=100 rather than 500, with a 1700 second timeout, and one step is one model request. Browser(local=True) now means Browser('chromium'). Browser('chrome') only attaches to a browser whose live DevToolsActivePort sits in your own profile directories, not to whatever answers on 127.0.0.1:9222. result.cost_usd covers model and web search charges but not the cloud browser or its proxy. is_successful() returns None when the agent gave no verdict. You can still pass your own model through ChatOpenAI, ChatAnthropic, ChatGoogle or a Pydantic AI model string.
The benchmark, read carefully
The post claims that on the same models the new version is up to 33 points higher, 2.7 to 5.2 times cheaper and 2 to 3 times faster. The published benchmark is more cautious. On BU Bench v2, 60 tasks run three times each on cloud browsers on 2026-11-09, browser-use 1.0 in ultrafast mode scored 62.5 against 56.3 for Browser Use Pi. An OpenAI outage hit that window: Pi lost 18 of 180 runs to provider errors and 1.0 lost one, which its single retry recovered from 13 of 14.
Excluding provider errors, the paired difference is +1.1 points with an interval of -3.2 to +5.7, which the release describes as no significant difference. Cost was $0.013 per task lower and time 114 seconds per task lower. Both arms ran gpt-6-luna on api.openai.com and were judged by gpt-5.6-luna at xhigh reasoning. The default bu-fast model on the Browser Use gateway was not part of the run.
Upgrading and what is new
MIGRATING.md lists every 0.x option and its 1.0 answer. Options 1.0 cannot honor raise a TypeError naming them before a browser starts or anything is billed, and removed imports raise an ImportError naming the replacement. Once configuration checks pass, run() does not raise: a rejected key, a browser that fails to start, or failures inside the run come back as status='error'.
New pieces include typed results through output_model_schema and result.structured_output, uploads and downloads through a workspace directory, secrets typed with type_secret() and redacted from cell output and results (the release is explicit that this is not a sandbox), domain restrictions, and web_search=True billed per search. Docker images are published at browseruse/browseruse:1.0.0 and ghcr.io/browser-use/browser-use:1.0.0, smoke-tested on amd64 and arm64. On reliability, Responses requests are streamed with 300 seconds to a first event and up to 930 seconds total, one transient retry per run, and compaction when a request exceeds the context window.
Limits and open questions
Every dependency is pinned exactly with a committed lockfile, which is a supply-chain decision. The cost is that a project needing a different version of pydantic, pydantic-ai-slim, mcp, google-genai or pillow cannot install 1.0 alongside it. google-genai stays below 3 because 3.0.0 calls the Gemini API on v1, which rejects tools. Viewport emulation does not work on cloud browsers yet. The release does not attach a methodology to the 33 point figure, and it does not say how the cloud browser or the gateway model are priced beyond per-search web search billing landing in cost_usd.
What it changes for app builders
If you are putting a browser step into an app, the surface is now smaller: a task string, optionally an output schema, and a structured result. The trade is that the default path runs on a hosted browser and a gateway key, so you are budgeting against Browser Use rather than your own model provider unless you opt into a local browser and your own model. If the app around the agent is the part you have not built, Draftship builds and hosts apps from a plain-language description with a built-in database and user sign-in, and its docs include a page on adding AI to your app. Draftship is not affiliated with browser-use.
Primary sources
- browser-use 1.0: CDP is all you need · Browser Use releases, Nov 10, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.