Browser Use 1.0 rebuilds web agents around Python and direct browser control

On November 10, 2026, Browser Use released version 1.0, changing how its AI agents interact with web pages.

The library moves away from DOM serialization and a fixed action schema. Instead, the model generates Python code executed in a persistent REPL, keeping variables, objects and state available across steps.

Browser operations use the Chrome DevTools Protocol (CDP). According to the release notes, the rewrite removes 136,866 lines of code, approximately 80% of the previous library. A smaller codebase does not, by itself, establish greater reliability.

An agent can combine Python instructions, persistent state and browser operations without selecting each step from a rigid set of commands. This increases flexibility while making controls over generated code, permissions and consequential actions more important.

Defaults have also changed: `Browser()` now uses Browser Use Cloud, while `Browser('chromium')` selects local execution. The default model is `bu-fast` through the Browser Use gateway, although other model providers can be configured.

The default step limit drops from 500 to 100, with a stated timeout of 1,700 seconds. The release includes structured outputs, uploads and downloads, secret redaction and domain restrictions. The migration guide documents compatibility changes for existing integrations.

In the BU Bench v2 results accompanying the release, 60 tasks were each run three times. Browser Use 1.0 ultrafast achieved a 62.5% success rate, compared with 56.3% for Browser Use Pi.

Reported average cost was $0.070 per task versus $0.082, and average duration was 650 seconds versus 743. These figures describe the tested configurations and conditions, not every deployment.

A provider outage affected the comparison. After excluding provider errors, the reported success-rate gap narrowed to 1.1 percentage points, with an interval from -3.2 to +5.7: not a statistically significant difference. The default `bu-fast` model was not the model evaluated in this comparison.

Browser agents can support enterprise workflows performed through web interfaces without suitable APIs: checking portals, retrieving information, filling forms and coordinating repetitive steps.

Dynamically generated code and direct browser access still require safeguards, including isolated sessions, least-privilege credentials, approval gates for irreversible actions and tests against changing page layouts.

Version 1.0 simplifies the framework architecture. Production suitability still depends on task-level reliability, error handling and the total cost of human oversight.