# Browser agent token statistics (2026, measured)

> Citable numbers from 19 real-website tasks in Claude Code: tokens, calls, cost, seconds and right answers for Playwright MCP, Playwright CLI and Dejo, what each turn costs, and site memory. Every number links its raw data.

Canonical: https://dejo.run/blog/browser-agent-token-statistics
Last updated: 2026-11-13

> For AI agents: the index of this site is https://dejo.run/llms.txt, and every page is also served as markdown (add .md to its address, or send Accept: text/markdown).

I build Dejo, so these are a tool maker's measurements, including of our own tool. Every number below links to the raw file it comes from, in the open-source [dejo-mcp repo](https://github.com/dejohq/dejo-mcp/tree/main/benchmarks). Quote them with a link, and say who measured them.

**Short version:** in fresh Claude Code sessions on 19 real-website tasks, a browser task cost **409K tokens** with Playwright MCP, **559K** with Playwright CLI and **251K** with Dejo the first time. Every turn re-sent about **39,700 tokens** before any page was read. Once a task was learned, Dejo repeated it for **89K tokens**, in a single tool call on 17 of 19 tasks.

## Tokens per task

Claude Sonnet 5, medium effort, one brand-new Claude Code session per task and tool, October 2026. Source: [`benchmark-real-2026-10-21.json`](https://github.com/dejohq/dejo-mcp/blob/main/benchmarks/benchmark-real-2026-10-21.json) and the Dejo reruns beside it.

| per task, 19 tasks | tokens | browser calls | cost | seconds | right answers |
|---|---|---|---|---|---|
| Playwright MCP | 409K | 6.5 | $0.233 | 40 | 17 of 19 |
| Playwright CLI | 559K | 8.4 | $0.276 | 38 | 18 of 19 |
| Dejo, first run | 251K | 4.7 | $0.173 | 31 | 19 of 19 |
| Dejo, every run after | 89K | 1.2 | $0.024 | 16 | 19 of 19 |

- The cheapest Playwright MCP task took 2 browser calls and 164K tokens; the most expensive took 20 calls and **1.36M tokens**.
- 6 more tasks were dropped because their sites turned away every headless browser (Cloudflare checks and plain 403 errors).

## Where the tokens go

From the same 19 Playwright MCP sessions:

- **39,713 to 39,739 tokens**: the first turn of every session, before any web page. Every turn re-sends it.
- **About 2,000 tokens**: the median growth of the context per turn after that.
- **12,777 characters**: the median size of a page snapshot (`browser_snapshot`, 32 calls). The largest was 47,491.

So the cost follows the number of turns, not page size. The full breakdown is in [why Playwright MCP uses so many tokens](https://dejo.run/blog/why-playwright-mcp-uses-so-many-tokens).

## Ten runs of the same task

| 10 runs of one task, averaged over 19 tasks | total tokens |
|---|---|
| Playwright CLI | 5.6M |
| Playwright MCP | 4.1M |
| Dejo (learn once, then replay) | 1.1M |

Runs 1 and 2 were measured; runs 3 to 10 repeat run 2's cost, because nothing changes for the tools without memory. Chart and method: [the benchmark](https://dejo.run/benchmark).

## A new task on a site the agent already knows

- **354K to 238K tokens (-33%)**: the average for a new task on 10 real websites, before and after Dejo had learned a different task on the same site. All 20 answers were right.
- **-64%** on rfc-editor.org, the biggest drop; **+28%** on pub.dev, where it cost more.

Source: [`benchmark-site-memory-2026-11-10.json`](https://github.com/dejohq/dejo-mcp/blob/main/benchmarks/benchmark-site-memory-2026-11-10.json). Write-up: [site memory](https://dejo.run/blog/browser-agent-site-memory).

## Finding the right element, no model involved

- **47 of 50 sites** for Dejo 0.4.0 and **45 of 50** for plain Playwright, on a reliability test with the same fixed steps for both: iframes, shadow DOM, custom dropdowns, date pickers, hover menus, infinite scroll and more.

Source: the `reliability-*.json` files and `reliability.py` in the [benchmarks folder](https://github.com/dejohq/dejo-mcp/tree/main/benchmarks).

To turn these into money for your own model and volume, use the [AI token cost calculator](https://dejo.run/tools/ai-token-cost-calculator) or the [Claude Code cost calculator](https://dejo.run/tools/claude-code-cost-calculator).

## What these numbers are not

- One model (Claude Sonnet 5) and one agent (Claude Code). Other models and agents will differ.
- Our tool and its alternatives, measured by us. The scripts are public so anyone can rerun them.
- Lesser-known websites on purpose: a model already knows famous sites' layouts, and sometimes their answers.


Related reading: [why Playwright MCP uses so many tokens](https://dejo.run/blog/why-playwright-mcp-uses-so-many-tokens), [site memory on a website the agent already knows](https://dejo.run/blog/browser-agent-site-memory) and [Playwright MCP vs Playwright CLI](https://dejo.run/blog/playwright-mcp-vs-cli).

## FAQ

## FAQ

### How many tokens does a browser agent use per task?

In our test of 19 tasks on 11 real websites with Claude Code (Claude Sonnet 5), Playwright MCP averaged 409K tokens per task, Playwright CLI 559K and Dejo 251K on the first run. Once Dejo had learned a task, each repeat averaged 89K.

### How much does a browser agent task cost?

The same runs reported an average of $0.233 per task with Playwright MCP, $0.276 with Playwright CLI and $0.173 for Dejo's first run. Dejo's repeat runs averaged $0.024.

### Can I quote these numbers?

Yes. Please link this page or the raw data file named next to each number. They are our own measurements of our own tool and its alternatives, and the method is public, so say that when you quote them.

## Related

- [Codex browser: what works in the CLI and IDE](https://dejo.run/blog/codex-browser)
- [Playwright MCP login: keep your agent signed in](https://dejo.run/blog/playwright-mcp-login)
- [Why Playwright MCP uses so many tokens](https://dejo.run/blog/why-playwright-mcp-uses-so-many-tokens)
- [Dejo MCP server setup: install in a minute](https://dejo.run/docs)
