A Guth Labs publication ↗
Back to News

Browser Use cuts reported model cost by preserving the prompt cache

Browser Use cuts reported model cost by preserving the prompt cache

Browser Use 1.0.4 changes screenshot cleanup so the browser agent does not continually invalidate its provider’s prompt cache. The maintainers report model cost per task of $0.208 versus $0.434 in a 200-task benchmark, with essentially unchanged scores, 76.1 versus 76.0. These are developer-reported results, not an independent replication. Release notes.

The mechanism matters more than the headline saving. Earlier cleanup repeatedly rewrote the conversation prefix. The new approach retains screenshots longer and removes old images together. The merged pull request explicitly identifies a tradeoff: the model can see more old screenshots, changing its context and potentially its behavior. Implementation discussion.

Guth analysis: Context management is an economic design decision, not merely housekeeping. Deleting material sooner may reduce the visible prompt while destroying useful cache reuse. A deployment evaluation should measure paid input, cache hits, task success and stale-screen mistakes together. The benchmark does not establish equivalent savings for every model, provider or workload.

Evidence limits: Both cited sources are project-maintainer sources. No independent test was performed. GitHub displays November 13 at 20:50 without an independently verified timezone. The PR confirms the November 13 merge date. This is newly discovered coverage of the November 13 release, not a release today.

Sources and citations

Revision history

Published .