Claude Code with a team vs Claude Code alone
Inside Herdr, Claude Opus 5.5 led a team: Codex (GPT-6.1 Sol) and Claude Sonnet 5.5, started and briefed by Claude itself in panes next to it. Against it, the same Claude Opus 5.5 working alone. Same brief, started in the same second, both recorded. Race 1 is one tight job (five files that depend on each other), race 2 is three games that share nothing. Everything here is exactly what the agents left: no fixes. It all runs in your browser.
A train-switch puzzle, 12 levels
Same brief, one agent
Orbit Relay · Lantern Maze · Stack Harbor
Same three games, one agent
The numbers
November 6, 2026, Windows 10, Herdr 0.9.5, on the Claude Pro and ChatGPT subscriptions (no API spend). Time = from the prompt to the lead's final report. Tokens = everything the agents read and wrote, from their own session logs; most of it is cached reads, which cost less. Checks = the same scripted real-browser test on both sides.
| Race | Side | Time | Tokens | Browser checks |
|---|---|---|---|---|
| 1 · Signal Yard | Team | 19 min 41 s | 11.47 M | 14 / 14 |
| 1 · Signal Yard | Solo | 19 min 7 s | 0.79 M | 14 / 14 |
| 2 · Arcade Shelf | Team | 12 min 40 s | 5.75 M | 33 / 33 |
| 2 · Arcade Shelf | Solo | 15 min 51 s | 2.70 M | 33 / 33 |
The exact prompts are for Ruocco Academy members. The Academy is not open yet: join the waiting list and you get one email when it opens. The demos on this page are free to use for everyone.
Join the waiting list