📅 November 9, 2026 • GitHub
browser-use, the open-source browser automation project, published release candidate 1.0.0rc4 on GitHub on November 9, 2026 as a pre-release. The repository now counts 120,445 stars and 13,459 forks, and this build is the first 1.0 candidate teams can actually pull and evaluate ahead of the stable tag.
The release replaces the engine (per Browser Use release notes). Instead of picking one action per step from a DOM dump, the agent now writes Python in one persistent REPL, with browser-harness helpers and raw CDP available in the same session, all built on Pydantic AI. The codebase shrank from 68,815 to 13,154 lines of Python, while the core 0.x API — Agent, Browser, Tools, run() and final_result() — is preserved.
New defaults change post-install behavior. llm=None now routes to the bu-fast model on the Browser Use gateway, and Browser() creates a Browser Use Cloud browser — pass Browser('chromium') for a local one. run() caps max_steps at 100 instead of 500 with a 1700 second timeout, a single BROWSER_USE_API_KEY replaces per-provider keys, and Python 3.11 or newer is required.
On BU Bench v2 — 60 tasks, 3 reps, cloud browsers — the 1.0 ultrafast mode scored 62.6 at 0.071 dollar per task and 585 seconds per task (per Browser Use). Against Browser Use Pi the score difference is reported as not significant, while the new build is 18 percent cheaper and 15 percent faster.
Version 1.0 also adds typed results through output_model_schema, so result.structured_output returns a validated object, plus file uploads and downloads, secret redaction via type_secret(), domain restrictions including block_ip_addresses, a web_search=True flag, an ultrafast mode driving accessibility-tree ids, and every dependency pinned exactly with == as a supply-chain policy.
Rollout note: rc1 through rc3 did not reach PyPI, so plain pip install browser-use still installs 0.13.14 until 1.0.0 ships. Teams that want the rewritten engine today need to install 1.0.0rc4 directly from the release tag.
For engineers, this moves browser agents from step-by-step action picking toward code-style control, with a far smaller codebase to audit and pinned dependencies for supply-chain review. For teams already on 0.x, the preserved API means existing Agent and run() code keeps working, while the new defaults route cloud usage through one key and tighter run limits.
Sources: GitHub browser-use releases, reveneau.com, whatsnew.fyi
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)