Browser Use 1.0.0 rebuilds its agent on CDP and runs 2 to 3 times faster

Browser Use 1.0 starts over. Your agent now writes Python in one persistent REPL, and every cell runs straight on Chrome DevTools Protocol, with nothing translating in between.

Slide text

+ 01

Browser Use 1.0.0 rebuilds its agent on raw CDP

The agent writes Python in one REPL, straight on Chrome DevTools Protocol

Release notes · browser-use 1.0.0

First, the new agent.

Slide text

+ 01

01

Release notes · browser-use 1.0.0

To get there, the team deleted eighty percent of the library. The DOM serializer and the action schema are gone. The model writes a cell, Chrome runs it, and the output goes back.

Slide text

+ 01

136,866 lines deleted, 80% of the library

● task
│
● model writes a cell    new_tab(url); bu.click(41)
│
● REPL runs it on Chrome (CDP), output back to the model
│
✓ finish(result)

Release notes · browser-use 1.0.0

On the same models, the new agent finishes far more tasks. Luna goes from passing about four tasks in ten to about seven. Every task costs a third as much or less, and takes less than half the time.

Slide text

+ 01

Same models, more tasks, a third of the cost

Release notes · browser-use 1.0.0 · benchmark chart

Next, the new defaults.

Slide text

+ 02

02

Release notes · browser-use 1.0.0

The default browser moved to the cloud. A plain browser with no arguments now starts on Browser Use Cloud. Tasks on localhost or a private network still need a local Chromium, and the agent warns you once.

Slide text

+ 02

Browser() now starts a Browser Use Cloud browser

Localhost and private IPs still need a local Chromium

Release notes · browser-use 1.0.0 · What changes by default

You no longer have to pick a model. Leave it empty and the agent runs bu-fast on the Browser Use gateway. So one Browser Use key replaces a separate key for every model provider.

Slide text

+ 02

Leave the model empty and it runs bu-fast

One Browser Use key replaces a key per provider

Release notes · browser-use 1.0.0 · What changes by default

Your own model still works. OpenAI, Anthropic and Google each have a chat class, and any Pydantic AI model string works too, Groq included. LangChain chat models are rejected now.

Slide text

+ 02

Your own model still works

• + Chat classes for OpenAI, Anthropic and Google
• + Any Pydantic AI model string, Groq included
• + LangChain chat models now raise a TypeError

Release notes · browser-use 1.0.0 · What changes by default

Old code fails loudly and early. An option the new agent can't honor raises a type error before a browser starts or anything is billed, and every removed import names what replaced it.

Slide text

+ 02

Old options fail before anything is billed

• + An option 1.0 can't honor raises a TypeError
• + A removed import names its replacement
• + Unused 0.x tuning knobs warn once

Release notes · browser-use 1.0.0 · Upgrading from 0.x

What else is new.

Slide text

+ 03

03

Release notes · browser-use 1.0.0

Hand the agent a Pydantic model and you get typed results back. It uploads and downloads files on local and cloud browsers, and it keeps secrets out of its output. That redaction is not a sandbox.

Slide text

+ 03

Typed results, files and secrets

• + Typed results from your own Pydantic model
• + Uploads and downloads on local and cloud browsers
• + Secrets redacted from output, though it is not a sandbox

Release notes · browser-use 1.0.0 · Also new

The short bu and browser commands are gone, because they broke uv tool installs beside other tools. The command is browser-use now. It pipes Python straight to browser harness and runs the MCP server.

Slide text

+ 03

The bu and browser commands are gone

• + browser-use pipes Python to browser-harness
• + It also runs the MCP server and installs the agent skill
• + browser-use init was removed

Release notes · browser-use 1.0.0 · Behavior changes · Also new

Long bu-smart turns no longer stop at five minutes. Every OpenAI Responses request now streams for up to fifteen minutes, and a failure hidden in a successful response gets one retry.

Slide text

+ 03

Long bu-smart turns no longer stop at 300 seconds

• + Every OpenAI Responses request streams, up to 930 s
• + A failure inside an HTTP 200 is retried once
• + An oversized request compacts once and is sent again

Release notes · browser-use 1.0.0 · Reliability

And who has to wait.

Slide text

+ 04

04

Release notes · browser-use 1.0.0

Every dependency is pinned to the exact version the tests ran. So a project that needs a different Pydantic, MCP or Pillow can't install Browser Use alongside it. The team wants to hear which pin blocks you.

Slide text

+ 04

1.0 pins every dependency exactly

A project that needs another pydantic, mcp or pillow version can't install it alongside

Release notes · browser-use 1.0.0 · Known issues

Cloud browsers can't emulate another screen size yet, so that still needs a local Chromium. And Google's gen AI library stays below version three until Pydantic AI supports it.

Slide text

+ 04

Cloud browsers can't change the viewport yet

• + Another viewport needs a local Chromium
• + google-genai stays below 3 until Pydantic AI supports it

Release notes · browser-use 1.0.0 · Known issues

I'd start any new browser agent on 1.0, because the same model finishes more tasks for a third of the cost. For older code, the migration guide maps every old option to its new answer, and the old versions stay installable.

Slide text

+ 04

I'd start any new browser agent on 1.0

Migration guide: MIGRATING.md

0.x stays installable

github.com/browser-use/browser-use

Release notes · MIGRATING.md · browser-use/browser-use