Your browser agent pays to look at pictures: opencli vs claude-in-chrome token use

Описание и применение

Сравнение токенов при использовании claude-in-chrome и opencli показало, что скриншоты стоят в 5 раз дороже текста. Opencli экономит ~27 % токенов, особенно на входных веб‑приложениях. Новое правило: использовать API, читать страницу как текст через opencli и делать скриншот только при необходимости визуальной проверки.

Задача

Determine token cost differences between Claude Code’s claude-in-chrome and opencli when interacting with web pages.

Результат

Screenshot calls cost ~5× more tokens than text calls; 23% of claude-in-chrome calls were screenshots, accounting for 49% of tokens. Opencli’s median tokens per call (397) matched claude-in-chrome text calls (373). Across nine sites, opencli saved ~27% tokens per call overall, especially on logged‑in web apps.

Рабочий процесс

• Collect 7,500+ browser tool calls from Claude Code logs.
• For each call, record API token counts before and after the call.
• Calculate token difference per call.
• Aggregate median tokens per call type for claude-in-chrome (screenshot vs text).
• Compare median tokens per call for opencli.
• Analyze token savings per site across nine websites with ≥10 calls each.
• Derive overall token savings and update clickops rule.

Доказательства

We went through our own Claude Code session logs: 7,500+ browser tool calls. For each call, we took the token counts the API reported before and after it. The difference is how many tokens that call's result added to the context.

Источник

A screenshot call cost about 5x a text call. The 23% of calls that returned screenshots made up 49% of all claude-in-chrome tokens.

Источник

opencli won on 7 of the 9 sites, about 27% fewer tokens per call when weighted by call count.

Источник

Our clickops rule now: 1) Use the API or curl when one exists. 2) Read the page as text with opencli. 3) Take a screenshot only to check something visual.

Источник