Tokens an agent reads
Loading the results…
snapshot --find, six sitesWhat is on screen, per site
The default snapshot: elements in the viewport plus an outline of the rest of the page. The dashed line is what one screenshot costs.
Every way of looking at a page
Tokens per call. Refs are the elements an agent can act on: the same count in both versions means nothing actionable was dropped.
| Site | Page text | Default 0.3 | Default 0.4 | Find 0.3 | Find 0.4 | Whole page 0.3 | Whole page 0.4 | Refs 0.3 | Refs 0.4 |
|---|
Whole tasks
Tokens the agent read to finish each task. No step failed in either version.
Step by step
| Task | Step | Tokens 0.3 | Tokens 0.4 | ms 0.3 | ms 0.4 |
|---|
Where 0.4.0 is not better
How it was measured
Both versions ran the same script back to back on the same machine: open each site, then record the tokens of the page as raw HTML, as visible text, as a screenshot, and as open-eyes' whole-page, default and --find snapshots. Three tasks then ran step by step. Tokens were counted with Anthropic's tokenizer; screenshots are estimated at width × height / 750. Live sites change between runs, so small differences on one page are noise.