Coding robots explained for tiny humans (and busy grown-ups). Tap things!
1
Three jobs, three helpers
Tap a layer to hear what it does.
2
Pick the hands 🦾
Same brain, different hands. How many jobs get done, and what does it cost? The numbers come from the Harness Leaderboard.
jobs
🎯 jobs done = jobs × the helper’s success rate. 🐷 piggy-bank = jobs done × cost per solved job. Gaps are small, so test on your own jobs before betting your lunch money.
3
Many robots? Pick a boss-board 🐑
Two free, open-source ones: T3 Code and Herdr. Answer, and watch the scores.
Grown-up fine print
Where the numbers come from
Leaderboard v0.1.9 (read 20 Nov 2026), all models and efforts, success rate averaged over Terminal-Bench 2.1 and DeepSWE. Cost uses list prices.
The brains tested are open models (GLM, DeepSeek, Flint, Ripple). So “Claude Code” here shows how it drives those, not how it drives Claude.
OpenCode: older write-ups show it far lower (35.5%, about $3.35 per solve). That was OpenCode 1.x. Version 2.2.6 replaced it on 9 Nov and sits at 52.5%.
Noise is big: re-running the same harness flips 24.1% of task outcomes, swapping the harness 27.1%, swapping the model 33.8%.
The board changes often. Check the live page before quoting a number.
About the boss-boards
T3 Code (MIT): lists Claude Code, Codex, Antigravity, OpenCode, Cursor and Grok, uses your own subscriptions, has a one-button commit, push and PR flow, plus iOS and Android apps. You can fork it to add an agent.
Herdr (Apache 2.0): runs real terminals, so anything that runs in a terminal works. Agents keep running when you close the window, machines over SSH share one sidebar, and one agent can start helper agents.
The step 3 scores are a simple fun rule of thumb, not a benchmark.