CLAIM-2026-08-13-04

Scaffold choice moves the result by roughly nineteen points for the same model: GPT-5.2 scores 21.8% under the mini-swe-agent scaffold and 41.2% under OpenHands. Both scaffolds were run with an identical step limit of 300 and a cost limit of $10 per instance. Every model except Gemini-3-Pro improves markedly on the same move.

Sources

• arxiv2026swebench