Source: github.com/browser-use/browser-use/releases/tag/1.0.4
The agent now removes old screenshots in one batch once 20 tool results contain one, rather than pruning each time a new screenshot arrives. The release notes say each earlier prune rewrote a prior message and broke the provider's prompt cache, so screenshot-heavy runs paid full input price on almost every step. In a 1.0.3 benchmark run, one 705-step task read 27.8 million input tokens, 1.6 million of them from cache.
Two smaller changes are included. Callers that set BU_NAME now get their own Chromium when Chrome cannot be used. When a browser setup error may be fixed in a newer release, the CLI and MCP tell the agent to upgrade.
On BU Bench v2.1, which has 200 tasks, 1.0.4 scored 76.1 against 76.0 for 1.0.3. Model cost per task fell from $0.434 to $0.208 over the same suite.
Why it matters if you build agents
If you run long, screenshot-heavy browser agents, upgrading should reduce input costs because cached prompt prefixes stay valid for longer. The cost figures come from one benchmark suite, so check them against your own traces before relying on them. Agents that hit browser setup errors may now receive upgrade guidance from the CLI and MCP.
Get stories like this every weekday
One short email with the day's agent tooling news.