Intel / repo

browser-use 1.0.4 — browser-use 1.0.4

Source: github.com
Author: gregpr07Top Viber
Date: Published Nov 13, 2026

Context

Version 1.0.4 of browser-use, an open-source Python library that lets an operate a web browser, roughly halves what its agent spends on model calls per task without lowering its score. On the project's own benchmark, BU Bench v2.1 with 200 tasks, model cost per task fell from $0.434 in version 1.0.3 to $0.208, while the score held at 76.1 against 76.0, according to the release notes. The saving comes from one change to when the agent removes old screenshots from its conversation history.

A browser agent sends its conversation so far to the model at every step, and screenshots make that history large. Model providers offer : when the beginning of a request matches a recent one, the matching part is read from a cache and billed below the full input price. That only works while earlier messages stay unchanged. Version 1.0.3 pruned an old screenshot each time a new one arrived, which rewrote an earlier message and broke the match, so screenshot-heavy runs paid full input price on almost every step. The notes cite one 705-step benchmark task on 1.0.3 that read 27.8 million input , only 1.6 million of them from cache. Version 1.0.4 waits until 20 tool results carry a screenshot and then drops the old ones in a single batch, so between those prunes the earlier history stays the same and can be read from cache.

Key takeaways · AI-distilled

browser-use 1.0.4 drops old screenshots in one batch once 20 tool results carry one, instead of pruning on every new screenshot, because each prune rewrote an earlier message and broke the provider's prompt cache.
The release notes cite a 705-step BU Bench 1.0.3 task that read 27.8M input tokens with only 1.6M served from cache, the cost that per-step pruning created on screenshot-heavy runs.
On BU Bench v2.1 (200 tasks), browser-use reports 1.0.4 scoring 76.1 against 76.0 for 1.0.3, while model cost per task fell from $0.434 to $0.208.
Callers that set BU_NAME now get their own Chromium when Chrome can't be used, and the CLI and tell the agent to upgrade when a browser setup error may be fixed in a newer release.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
prompt caching — Reusing the model's processed form of a repeated prompt prefix so subsequent calls skip re-reading it, cutting cost and latency substantially.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.

Read the source github.com