All capabilities Agents, tools and MCP

Drive a real browser, not a scraper

The model gets a live browser: it navigates, reads the accessibility tree, clicks by element reference, fills forms and manages tabs, in a loop, until the task is done.

DocsOfficially documentedTestNot tested by benchr Medium · API Docs checked September 1, 2026

Claude APIClaude CodeClaude Fable 5Claude Opus 5+2

Overview

Exposes 31 member tools under one toolset - navigate, screenshot, read_page, find, form_input, tab management, pointer and keyboard actions.
Returns an accessibility tree with element references, so a click targets an element instead of a pixel guess.
Runs as an agent loop: the model emits member calls, your executor runs them, results go back, and it continues.

How it works

01Add the toolset to the request as browser_toolset_20260801 on a supported model.
02Run the browser in an isolated container or VM behind a network-level domain allowlist. The documentation states this as a requirement, not a suggestion.
03Prefer read_page and find over screenshots when verifying text and structure - references are cheaper and less ambiguous than coordinates.
04Leave javascript_exec, file_upload, read_console and read_network disabled unless the task genuinely needs them. All four ship off by default for reasons the docs spell out.
05Require human confirmation before consequential actions such as purchases, account changes or sending messages.

Limits

Element references are scoped to their tab and go stale as soon as the tab navigates or the DOM changes materially.
Coordinates are viewport pixels only - no desktop or window frame - and screenshots are not automatically downscaled to the model's image limit.
read_page output is capped at 50,000 characters.
Batched member calls stop at the first failure, and every remaining call must still be answered.
Not available on Amazon Bedrock, Claude Platform on AWS or Microsoft Foundry. Page content, tab titles and URLs are named prompt-injection surfaces.

Verification

benchr test runs

benchr has not run this capability. No test has been published. 1 self-reported run awaiting independent review.

benchr test runs →

Evidence

Sources

Anthropic - Browser use tool (browser_toolset_20260801) Tier 1 · September 1, 2026

Not stated by the source

Success rates on any particular website
Per-action latency or cost beyond standard token pricing

Source-check history

September 1, 2026DocsOfficially documentedFirst logged against browser_toolset_20260801, with the model list, the 31 member tools and the disabled-by-default set read directly from the tool page.

Changes

Provider changes affecting this

August 31, 2026Claude Sonnet 5Pricing changed
July 3, 2026Claude Sonnet 5Pricing changed
July 1, 2026Claude Sonnet 5Model released
June 10, 2026Claude Fable 5Model released