Stagehand — the SDK for browser agentsEvalsBlogShowcase

// Research & public data

Turn a leaderboard into structured data

Read two benchmark leaderboards on stagehand.dev/evals and return the top 10 models of each as JSON.

stagehand.dev ↗
stagehand code-mode agent (mcp facade), browserbase recording at 2× speed, no sound

5 agent tool calls, click to jump

++

Why it’s hard

The data lives in an interactive client-side view, not a file you can download.

  • Each benchmark sits behind a switcher that re-renders the table in place
  • Rows mix percentages, dollars and seconds in display formatting
  • Charts and tables show the same models in different orders

Stagehand vs. Playwright MCP

Given the same goal and model, the Stagehand code-mode agent finished for 52% less than an agent driving Playwright MCP.

stagehand code mode agent3/3 runs succeeded
$0.046
playwright mcp agent3/3 runs succeeded
$0.095

medians over successful runs, or over all runs when none succeeded; lower is better

Then run it as a script

Once the flow works, the cookbook runs it as plain Stagehand calls. With caching on, a re-run cost 100% less, because Stagehand served unchanged steps from its cache.

cookbook, first run3/3 runs succeeded
$0.096
cookbook, cached re-run3/3 runs succeeded
$0

medians over successful runs, or over all runs when none succeeded; lower is better

anthropic/claude-sonnet-5 at $2 input · $0.2 cached · $2.5 cache write · $10 output per 1m tokens · stagehand 4.1.0 · playwright mcp 0.0.82 · agents stop after 60 llm calls · run on 2026-09-28 · methodology

++

Run it yourself

Give the same goal to Claude Code with Stagehand’s code-mode MCP server, which exposes three tools: run, snapshot and screenshot. Once the flow works, the cookbook runs it as a plain Stagehand script.

cookbook on githubcode-mode setup on github
the goal both agents were given

On stagehand.dev/evals, read the leaderboard for the "Browserbase Benchmark v2" benchmark and then for the "Online Mind2Web" benchmark. For each, report the top 10 rows with rank, model, harness, accuracy in percent, cost per task in USD and seconds per task (null if not shown).

++

More in research & public data

see all →
++