What changed
Browser Use released 1.0.4 with batched screenshot pruning after 20 screenshot-bearing tool results. The maintainer says the earlier per-step rewrite invalidated provider prompt caches; its 200-task BU Bench run held score roughly flat while reported model cost per task fell from $0.434 to $0.208.
Why it matters
Browser agents can silently multiply input-token cost when their own history compaction rewrites cached prefixes. The release identifies a concrete runtime cause and makes history mutation, cache hits and per-task spend observable deployment checks.
What to do
Pin 1.0.4 in a sandbox and replay a screenshot-heavy task with fixed model, browser state and seed. Record input, cached and output tokens, task score, pruning events and wall time against 1.0.3 before changing production limits.
Caveat
The cost and score figures are maintainer-run results on BU Bench, not independent evidence or a universal saving. Gains depend on provider caching and screenshot volume; the release still pins Pydantic AI dependencies that can conflict with other agent packages.