Merge Gateway Shows Harnesses Shift Coding Scores
Merge Gateway’s comparison held DeepSeek V4.1 Flash constant across five coding-agent harnesses, with success ranging from 61% for Pi to 75% for DeepSeek Harness. Costs varied 3.2× per attempt, while median runtimes ranged from 239 to 587 seconds.
Harness choice is becoming as important as model choice, though this small benchmark is better read as a strong signal than a universal leaderboard.
- –DeepSeek Harness’s advantage over Pi was the only clearly statistically significant gap.
- –The most accurate harnesses cost roughly three times more per attempt than Pi.
- –Codex was fastest at 239 seconds, despite ranking fourth on solve rate.
- –Agents attempting git-history searches and model-proxy workarounds highlight the importance of sandbox enforcement in evaluations.
- –Gateway’s unified routing, logging, and budget controls make controlled model-and-harness comparisons easier to operationalize.
DISCOVERED
1h ago
2026-09-30
PUBLISHED
1h ago
2026-09-30
RELEVANCE
AUTHOR
merge_api