CLAIM-2026-08-13-06

Open-weight models come within a few points of the frontier at a fraction of the cost: GLM-5 and Qwen3.5 both score 36.5% and Kimi-K2.5 32.9%, at $0.24, $0.78 and $0.72 per instance respectively, against GPT-5.2 at 41.2% for $3.60 and Claude Sonnet 4.6 at 38.8% for $4.77.

Sources

• arxiv2026swebench