Browser Use 1.0.4 cuts token costs by batching screenshot cleanup
The main change is how old screenshots are pruned: instead of dropping one on every new screenshot, the agent now removes them in a single batch after 20 tool results carry an image. Because each earlier prune rewrote a prior message and invalidated the model provider's prompt cache, screenshot-heavy runs were paying full input price on nearly every step. The release notes cite a 705-step task that read 27.8M input tokens, only 1.6M of them cached. Also included: `BU_NAME` callers get their own Chromium when Chrome can't be used, and the CLI and MCP now suggest upgrading when a browser setup error may be fixed in a newer release.
Why it matters
Anyone running long, screenshot-heavy browser agent tasks can spend roughly half as much: Browser Use reports $0.208 model cost per task on BU Bench v2.1 versus $0.434 on 1.0.3, with the same score.
Sources
- PrimaryBrowser Use
Written by AIToolMaps from the source above and checked against it sentence by sentence.