Browser Use 1.0: New Agent Engine, Defaults and Migration

Q: What is new in Browser Use 1.0?

The agent no longer picks one action per step from a DOM dump. It writes Python in one persistent REPL that runs directly on Chrome DevTools Protocol. About 80% of the library was removed, and the defaults changed to the bu-fast model, a Browser Use Cloud browser and a single Browser Use API key.

Q: Can I still use my own model and a local browser?

Yes. Pass ChatOpenAI, ChatAnthropic, ChatGoogle or a Pydantic AI model string as the model. Use Browser('chromium') for a local Chromium, or Browser('chrome') to attach to your own open Chrome.

Q: Will my 0.x code break?

Possibly. Common entry points such as Agent, Browser, run() and final_result() still work, but removed imports raise an ImportError naming the replacement, and options 1.0 cannot honor raise a TypeError before anything is billed. If you need 0.x, install browser-use<1.

Q: Which Python version does Browser Use 1.0 need?

PyPI lists Python 3.11 or newer (below 4.0). The package is released under the MIT license.

Q: Is Browser Use 1.0 better than 0.x?

The project reports up to 33 points higher scores with the same models, 2.7 to 5.2 times lower cost and 2 to 3 times faster runs. Its own BU Bench v2 comparison against Browser Use Pi found no significant quality difference once provider errors were excluded, but lower cost and time per task.

Last updated: November 10, 2026

Browser Use 1.0, released November 10, 2026, replaces the open-source browser agent’s engine. The agent now writes Python in one persistent REPL that drives Chrome through the DevTools Protocol. Defaults move to the bu-fast model and a cloud browser, and the library is about 80% smaller.

What’s new in 1.0
0.x vs 1.0 defaults
How to install and run Browser Use 1.0
New features for power users
Benchmark results
Migrating from 0.x
Known issues and limitations
Frequently asked questions
Sources

What’s new in 1.0

The 1.0.0 release notes describe a rewrite: 136,866 lines removed and 80% of the library deleted. In each step, the model writes a Python cell, such as opening a tab or clicking an element, and the REPL runs it on Chrome DevTools Protocol (CDP). The output goes back to the model until it calls finish(result).

There is no DOM serializer and no fixed action schema in between. The project states that, with the same models, 1.0 scores up to 33 points higher, costs 2.7 to 5.2 times less and runs 2 to 3 times faster than 0.x.

A follow-up 1.0.1 release later the same day fixed MCP use on machines without Chrome, Windows UTF-8 output, timeout handling and per-run workspace isolation.

0.x vs 1.0 defaults

The biggest practical change is what happens when you do not pass a model or browser.

Setting	0.x	1.0
Model (`llm=None`)	Required	`bu-fast` on the Browser Use gateway
`Browser()`	Local Chromium	A Browser Use Cloud browser; `Browser('chromium')` for local
API key	One per model provider	One Browser Use key (`BROWSER_USE_API_KEY` or `browser-use login`)
Mode	Not applicable	`ultrafast` (accessibility-tree IDs) on bu-fast and bu-ultrafast; `mode='default'` opts out
`run()` limits	`max_steps=500`	`max_steps=100`, `timeout=1700` seconds

One step equals one model request, so the lower step cap also limits spend on runaway tasks.

How to install and run Browser Use 1.0

Confirm you are on Python 3.11 or newer, as required on PyPI.
Install with an exact pin, as the project recommends: uv add 'browser-use==1.0.1' (or pip install 'browser-use==1.0.1').
Create a key at cloud.browser-use.com and export it: export BROWSER_USE_API_KEY=bu_.... You can also run browser-use login.
Run a first task: from browser_use import Agent, then Agent('Find the number of stars of the browser-use repo').run_sync().final_result().
For sites on localhost or private IP addresses, pass Browser('chromium'), because the default cloud browser cannot reach them.
To use your own provider, pass a model such as ChatAnthropic or a Pydantic AI string like llm='groq:llama-3.3-70b-versatile'.

New features for power users

Typed results: pass output_model_schema=MyModel to Agent and read result.structured_output.
Uploads and downloads: inputs come from workspace=; downloads land in workspace/downloads and result.downloads, on local and cloud browsers.
Secrets: sensitive_data= values typed with type_secret() are redacted from cell output, checkpoints and results. The notes stress this is not a sandbox.
Domain rules: allowed_domains, prohibited_domains and block_ip_addresses work in every browser mode. A prohibited domain now also blocks its subdomains.
Web search: web_search=True is billed per search to your Browser Use key and included in result.cost_usd.
CLI and MCP: browser-use --cli-mcp runs the MCP server, and browser-use skill install registers the agent skill. Since 1.0.1, the MCP server installs its own Chromium on servers and CI machines without Chrome.
Docker: images are published as browseruse/browseruse:1.0.0 and ghcr.io/browser-use/browser-use:1.0.0.

Reliability changes include streamed OpenAI Responses requests (up to 930 seconds per turn), one automatic retry for transient provider failures, and automatic context compaction when a request exceeds the model’s window.

A practical prompt pattern from the notes: to collect partial results on long jobs, name the checkpoint in the task, for example “after each page, save the titles with checkpoint(‘titles.json’, titles, partial=True)”.

Benchmark results

The release reports BU Bench v2 (60 tasks, 3 repetitions, cloud browsers, November 9, 2026). Browser Use 1.0 in ultrafast mode scored 62.5 at $0.070 and 650 seconds per task, versus 56.3, $0.082 and 743 seconds for Browser Use Pi. The project notes that an OpenAI outage inflated the gap. With provider errors excluded, quality showed no significant difference, while 1.0 was cheaper and faster. Both arms used gpt-6-luna, not the default bu-fast model.

For comparison with model-native approaches, see our guides to Gemini Computer Use, the open-weight NorthCUA computer-use models and the Mage-VL-Cowork GUI agent.

Migrating from 0.x

Read MIGRATING.md first; it maps every 0.x option to its 1.0 answer. Key behavior changes:

Removed imports raise an ImportError naming the replacement. Options 1.0 cannot honor raise a TypeError before any browser starts or anything is billed.
After configuration checks pass, run() does not raise. Failures return status='error' with result.error.
is_successful() returns None when the agent gave no verdict.
result.cost_usd covers model and web search charges, not the cloud browser or its proxy.
LangChain chat models now raise a TypeError, and the 0.x gateway models such as bu-2-0 are rejected.
The bu and browser console commands are gone; use browser-use.

To stay on the old engine, pin pip install 'browser-use<1'.

Known issues and limitations

Every dependency is pinned exactly, so projects that need a different version of packages such as pydantic, mcp or google-genai cannot install 1.0 alongside them. The google-genai package stays below version 3. Viewport emulation does not yet work on cloud browsers. The release notes do not list prices for bu-fast or cloud browsers.

If you only need search results inside an agent rather than full browsing, our AgentCore Web Search API guide and OpenAI Agents API guide cover managed alternatives.

Frequently asked questions

What is new in Browser Use 1.0?

The agent no longer picks one action per step from a DOM dump. It writes Python in one persistent REPL that runs directly on Chrome DevTools Protocol. About 80% of the library was removed, and the defaults changed to the bu-fast model, a Browser Use Cloud browser and a single Browser Use API key.

Can I still use my own model and a local browser?

Yes. Pass ChatOpenAI, ChatAnthropic, ChatGoogle or a Pydantic AI model string as the model. Use Browser(‘chromium’) for a local Chromium, or Browser(‘chrome’) to attach to your own open Chrome.

Will my 0.x code break?

Possibly. Common entry points such as Agent, Browser, run() and final_result() still work, but removed imports raise an ImportError naming the replacement, and options 1.0 cannot honor raise a TypeError before anything is billed. If you need 0.x, install browser-use<1.

Which Python version does Browser Use 1.0 need?

PyPI lists Python 3.11 or newer (below 4.0). The package is released under the MIT license.

Is Browser Use 1.0 better than 0.x?