Skip to content

browser-use 1.0.4 Fixes Prompt Cache Breaks to Cut Run Cost

The release batches screenshot pruning so prompt caches survive, and the benchmark cost per task drops from $0.434 to $0.208.

Draftship Editorial2 min read

A caching fix, not a capability fix

On 13 November, Browser Use published browser-use 1.0.4, installed with pip install 'browser-use==1.0.4'. The library drives a browser with an AI agent, and the release is mainly about what those long runs cost. The agent now discards old screenshots in a single batch once 20 tool results carry one, rather than doing it every time a new screenshot arrives.

The source is clear that this is a cost change, not a capability change. On BU Bench v2.1, a 200 task benchmark, the score was essentially flat against 1.0.3 (76.1 versus 76.0), while model cost per task fell from $0.434 to $0.208.

Why pruning every screenshot was expensive

The mechanism is prompt caching. Providers cache the repeated front part of a conversation, so steps that reuse it are billed at a lower input rate. Every time the agent pruned a screenshot, it rewrote an earlier message. That invalidated the cached prefix, so the next step paid full input price for the whole history again.

In screenshot-heavy runs, that happened on nearly every step. The release gives a BU Bench figure from 1.0.3 as the illustration: a single 705 step task read 27.8M input tokens, of which only 1.6M came from cache.

Batching the clean-up means the earlier part of the conversation stays untouched between prunes, so the cache can be reused across steps. The agent still loses old screenshots, it just does so less often and all at once.

Two smaller changes in the same release

The release notes list two other changes, with far less detail than the caching work. Callers using BU_NAME get their own Chromium when Chrome cannot be used (#6387). And when a browser setup error looks like it may be fixed in a newer version, the CLI and the MCP integration now tell the agent to upgrade (#6388). The source does not say what BU_NAME is beyond that name.

Limits and open questions

The notes do not say which model or provider the cost figures were measured against, so the $0.208 figure should be read as one benchmark result rather than a general price. The 20 result threshold is stated flatly, with no indication that it is configurable. The source also does not describe what internal caching assumptions the fix relies on, or whether every provider caches conversation prefixes the same way. There is no breaking-change note, which suggests existing callers can upgrade without code edits, but that is an inference rather than something the release states.

What it means for building apps with AI

The lesson travels beyond this library. Agent loops that rewrite earlier messages quietly destroy cache performance, and in screenshot or document-heavy apps that can dominate the bill. If you build AI features into an app, it is worth checking how your history is assembled between steps. Draftship's documentation on adding AI to your app covers where model calls sit in an app you build there. Draftship is not affiliated with Browser Use and does not support anything described in this release.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.

AI coding agents

opencode v2.3.2 fixes GPT-6.2 prompt caching billing

opencode v2.3.2 sends the system prompt as a leading developer message on every Responses route so GPT-6.2 caches it, cutting uncached input per step from about 14,000 tokens to 3 on one Copilot model.