CLAIM-2026-08-13-08

OpenAI published 'Why SWE-bench Verified no longer measures frontier coding capabilities' on 23 February 2026. It audited 138 failed problems and found 59.4% contained material issues in test design: 35.5% narrow tests enforcing specific implementation details that invalidate functionally correct submissions, and 18.8% wide tests checking functionality the problem description never specified. NOTE: this is a February post, not a this-week event; several aggregators resurfaced it during the week of 10 August 2026.

Sources

• openai2026why