the top AI browser agents score about the same on live websites, so speed and then cost decide how much work you get done.

five setups ran the same Online-Mind2Web tasks on live sites: OpenClaw browser, Hermes browser, Claude in Chrome, 5dive browser (Claude Code) and 5dive browser (pi). pi is a minimal coding agent. all on Claude Haiku 5.5, with the same prompts and a 10-minute cap.

100 live websites

• OpenClaw browser finished 35, 5dive browser (Claude Code) 32, Claude in Chrome 29. none of those gaps is statistically significant (paired p from 0.26 to 0.71).
• median time per task: Claude in Chrome 67 seconds, Hermes browser 86, 5dive browser 105 in both agents, OpenClaw browser 141.
• a faster 5dive browser is next on our roadmap.
• the benchmark is hard on purpose. its paper reports many agents finish about 30%.
• cost per finished task: $0.04 for 5dive browser (Claude Code), $0.02 for 5dive browser (pi), about $0.03 for Claude in Chrome, $0.14 for OpenClaw browser.
• Hermes browser finished 18. headless, its default, it finished 13 and hit 39 bot walls.
• with a window on every setup, bot walls ranged from 20 to 26 of 100.

4 everyday tasks

three form edits on a test app we built and one Wikipedia lookup, 3 runs each. every agent passed all 12 runs.

5dive browser (pi) did them for $0.86 per 1,000 tasks. OpenClaw browser cost $6.92, eight times more. steps here count model calls. the first picture counts tool calls.

why the costs differ

one call to the 5dive browser fills a form, clicks and checks the result. OpenClaw browser does one action per call.

most of the cost gap is the agent around the browser. every call re-reads its prompt: about 73K tokens in OpenClaw, 36K in Claude Code, 4.5K in pi. on the everyday tasks, the same 5dive browser cost about a quarter as much in pi as in Claude Code.

on a server

a signed-in browser job on a schedule took the 5dive browser 4 steps and nothing built by hand. OpenClaw and Hermes browsers needed services we built by hand and a borrowed remote desktop to sign in.

Claude in Chrome clicks in your own desktop Chrome, with your logins. to run it on a server, we had to hand-build a screen and a remote viewer.

on 28 October an agent with the 5dive browser opened a supplier account, made its API key and built a Meta ad up to the Publish button. the owner did one Google tap and one Facebook sign-in.

the agent never types your password. when a site throws a robot check, it hands you the live window.

limits

• one run per task, one model, one server, one datacenter IP that some sites reject.
• we built the test rig. WebJudge on Claude Haiku 5.5 scored it, not the benchmark’s own judge, so don’t compare with the public leaderboard.
• costs are API list prices for the logged model calls, not what a flat subscription costs. Claude in Chrome runs on a subscription, so its cost is an estimate.
• Claude in Chrome ran a day after the others, on live sites that may have changed.
• OpenClaw can also batch actions. we didn’t test that.

run it yourself

the 5dive browser is free and Apache 2.0 licensed.

curl -fsSL https://install.5dive.ai | sudo bash
sudo 5dive plugin add 5dive-ai/5dive-browser
sudo 5dive browser setup
sudo 5dive init   # your first agent, after setup so it gets the browser

---

use github.com/5dive-ai/5dive to self-host.

use 5dive.ai for a managed box.