“A qualitative evaluation by DataCurve found that self-verification was the biggest differentiator on the DeepSWE benchmark, with GPT-5.4 and Opus 4.7 writing their own tests to verify work over 80% of the time.”