Benchmarks
Coding task results
Four coding tasks, five attempts per task. Handwork was rerun on October 19, 2026. Other agents use earlier results.
| Agent | Complete fixes | Median time | Tool calls | Input tokens | Output tokens | Peak agent RSS |
|---|---|---|---|---|---|---|
 handwork ReleaseFast | 5 / 5 | 29.0 s | 5.0 | 30,741 | 1,012 | 24.4 MiB |
 OpenCode | 5 / 5 | 55.2 s | 15.8 | 78,781 | 1,139 | 790.1 MiB |
| Codex | 5 / 5 | 32.8 s | 4.2 | 82,758 | 763 | 234.0 MiB |
| Claude Code | 5 / 5 | 21.2 s | 3.0 | 37,797 | 2,096 | 275.2 MiB |
Fix bars show pass rates. Other bars scale to the largest value in each column. Tokens and tool calls are means per attempt. Bar length shows magnitude, not an overall score.
We kept every result, including failures. OpenCode and Codex ran on September 18-19, and Claude Code on October 4, 2026. The older Handwork results remain in the archived reports. Cache results use a separate complete five-attempt rerun on October 19, 2026.
Handwork, OpenCode, and Codex used GPT-6 Astra at low effort. Claude Code used Opus 5.5 at medium effort. Different models and run dates limit the comparison. These four tasks do not establish an overall ranking.