Same model, two harnesses: 24 of 25 versus 19 of 25 on merged PRs

What the research found

Octomind's octobench ran four coding agents on 25 tasks from 2026 merged PRs across 5 languages; octomind with glm-5.2 solved 24/25 versus opencode with glm-5.2 at 19/25, showing the harness alone changes results on the same model.

Claims and how well the evidence supports them

1. The octomind plus glm-5.2 harness solved 24 of 25 tasks, while opencode plus glm-5.2 solved 19 of 25 95%
2. All 25 tasks had fixes merged in 2026 and were validated fail-to-pass against held-out project tests 88%
3. Claude Code with claude-opus-5 solved 23 of 25 tasks at a cost of $81.79 93%
4. Codex with gpt-5.6-sol solved 21 of 25 tasks at a cost of $14.86 95%

Evidence

• octomind.run https://octomind.run/blog/coding-agent-benchmark-real-prs

Actionability rated low.