Browser Use Deleted 80% of Its Library and Scored Up to 33 Points Higher
On November 10, 2026, browser-use shipped version 1.0 under a release title that reads like a taunt: CDP is all you need. The notes lead with a deletion statistic rather than a feature list. 136,866 lines removed, roughly 80 percent of the library. News coverage put the surviving Python core at about 13,700 lines, down from nearly 69,000.
This was not a cleanup. The engine underneath a familiar public API got replaced. The Agent, Browser, run() and final_result() surface still works, which is why the rewrite deserves a close read: the same interface now hides a completely different theory of what a browser agent should be.
The old engine, and why they gave up on it
In 0.x, the library serialized the page into an indexed DOM dump, handed it to the model, and the model picked one action from a fixed schema: click(index=12), type this, scroll that. Every capability had to exist as a predefined action before the model could reach it. Looping over ten search results meant ten round trips, each paying for a fresh DOM dump. That design assumed the model was fragile and needed guardrails. The new one assumes the opposite.
What 1.0 does instead
The 1.0 agent writes Python. Each step, the model emits one cell, something like new_tab(url); bu.click(41); js(...). A persistent REPL runs it against Chrome over the Chrome DevTools Protocol, the same interface the browser's own developer tools use, with browser-harness helpers plus raw CDP calls. The output comes back to the model, which writes the next cell, until it calls finish(result). No DOM serializer. No action schema. Nothing in between.
The loop runs on Pydantic AI, so any Pydantic AI model works and you inherit its OpenTelemetry traces. The architecture is a bet on what current models are good at: Python supplies control flow and state, CDP supplies browser access. A single cell can click, type, read, and loop. Anything Chrome DevTools can do, the agent can do, without anyone shipping a built-in action for it first. In the action-schema world, every feature request became a new action, a new schema field, new tests. In the code world, the model writes the feature.
What the benchmark actually shows
The release reports BU Bench v2 (60 tasks, 3 repetitions, cloud browsers, run 2026-11-09) with 1.0 in its default ultrafast mode, which addresses elements by accessibility-tree ids:
| Agent | Score [95% CI] | $/task | s/task |
|---|---|---|---|
| browser-use 1.0 (ultrafast) | 62.5 [58.3, 66.5] | $0.070 | 650 |
| Browser Use Pi (previous agent) | 56.3 | $0.082 | 743 |
Against 0.x with the same models, the release claims gains of up to 33 points, costs 2.7 to 5.2 times lower, and runs 2 to 3 times faster. Hold those numbers loosely: the benchmark ran on the 1.0.0rc2 build, the judge is the company's own findings rubric, and early coverage noted the gap over Pi is not statistically significant once provider errors are excluded. The direction is credible. The precision is not.
The defaults quietly point at a paid gateway
If you leave everything unset, 1.0 no longer runs on your machine or your keys. llm=None now resolves to bu-fast on the Browser Use gateway instead of raising (it was a required argument in 0.x), and Browser() starts a Browser Use Cloud browser, stopped when the run ends. One BROWSER_USE_API_KEY replaces one key per model provider. A local browser needs Browser('chromium'), and tasks on localhost or private IPs need one, with a warning that says so. The per-run limits also moved: max_steps defaults to 100 instead of 500, with a 1,700-second timeout.
Then the ground moved again. Two patch releases shipped within two days of launch, and 1.0.6 (November 16) made BU 4.0 the default model, which costs 15 times as much as bu-fast ($1.80/$9.00 versus $0.12/$0.60 per million input/output tokens). The release notes' own test task went from $0.0014 to $0.022. If you upgraded and pass no llm, you paid it. The fix is one line (llm='bu-fast'), but only if you notice. The project's README now tells you to pin with ==, never >=, because releases move fast and can break APIs. With defaults this mobile, that is not boilerplate advice.
The migration traps that bite in production
MIGRATING.md lists every removed option with its 1.0 answer, and removed features raise clear errors naming the replacement before a browser starts or anything bills. The traps are in the quieter changes:
- Failures stop raising. Once configuration checks pass,
run()returnsstatus='error'withresult.errorinstead of throwing. Code that wraps runs intry/exceptwill sail past failures. Checkresult.errorafter every run. - The success verdict changed shape.
is_successful()returnsNonewhen the agent gave no verdict. The 0.x behavior maps tor.is_done()andr.success is not False. - Error history is split.
errors()returns only the error that ended the run. Recovered cell errors live inresult.messages. - Attaching to your own Chrome got stricter.
Browser('chrome')attaches only to a browser whose liveDevToolsActivePortfile sits in your own profile directories. It no longer grabs whatever answers on127.0.0.1:9222, which is the right call for security and a surprise for anyone with a shared debugging setup. - Dependency pins collide. 1.0 pins
pydantic,mcpandgoogle-genaiexactly, so it can fail to install beside projects that pin those differently. Resolve your dependency set in a branch first, or stay on 0.x withpip install 'browser-use<1', which remains installable and documented.
The prompt-cache bug teaches more than the benchmark
The most transferable bug report in this release train is in the 1.0.3 notes. The agent prunes old screenshots from history to save tokens, but it pruned one at a time, and every prune rewrote an earlier message. That breaks the provider's prompt cache from the rewritten point forward, so screenshot-heavy runs paid full input price on almost every step. One 705-step task read 27.8 million input tokens with only 1.6 million served from cache. The fix batches the drop until 20 tool results carry screenshots, and on BU Bench v2.1 (200 tasks) the score held at 76.1 versus 76.0 while model cost per task fell from $0.434 to $0.208.
Should you move
New project: start on 1.0, pin the version exactly, choose your llm explicitly instead of trusting the default, and decide between cloud and local browsers before writing tasks that touch localhost.
Existing 0.x in production: nothing forces the upgrade. Re-check your failure handling and verdict logic against the changes above, test the dependency resolution, and rerun your own task set rather than trusting BU Bench, whose judge scores against the vendor's rubric.
What I find interesting is the sociology of it. For two years the standard recipe for a browser agent was serialize the page, constrain the actions, keep the model on a short leash. Browser Use was the most serious implementation of that recipe, and its own team decided the scaffolding had become the bottleneck once the models could write decent Python. The deleted 80 percent was mostly the part that existed because the model couldn't be trusted. The deletion is the result.
Comments