GPT-5.3-Codex stalls on debugging benchmark

This is the kind of benchmark developers should care about more than leaderboard candy, because diagnosing ugly real-world failures is closer to senior engineering work than solving curated coding tasks.

• –OpenAI-Proof Q&A is based on internal bottlenecks that reportedly cost OpenAI teams at least a day each, so it measures practical research friction rather than toy problems
• –GPT-5.3-Codex looks strong on other long-horizon and cyber evaluations, which makes the flat result here a useful reminder that “better at coding” does not automatically mean “better at debugging”
• –If models plateau on root-cause analysis of messy training and systems failures, the timeline to fully autonomous AI research engineers still looks longer than hype suggests
• –The caveat is that this is an internal, unpublished eval, so outside researchers cannot independently validate task design or grading yet