OpenAI's flagship surpasses all previous models on the canonical software-engineering benchmark.
GPT-5 reclaimed the top spot on SWE-bench Verified this morning at 78.4%, edging Claude 4.6 Sonnet's previous high of 76.1%.
SWE-bench measures end-to-end issue resolution on real Python repositories. The benchmark is widely seen as the most realistic proxy for production coding-agent quality, though critics note its bias toward Python web tooling.
OpenAI attributed the gain to a new tool-use post-training recipe rather than a base model swap.
SWE-bench Verified is a curated subset of SWE-bench that measures end-to-end issue resolution on real Python repositories, widely treated as the most realistic proxy for production coding-agent quality.
OpenAI attributed the 2.3-point gain (78.4% vs 76.1%) to a new tool-use post-training recipe rather than a base model swap.
Critics note the benchmark's bias toward Python web tooling; it under-samples systems languages, frontend work, and large-scale refactoring tasks, so leadership on SWE-bench does not perfectly predict day-to-day developer experience.