GPT-5.2

GPT-5.2 sets new benchmarks in reasoning across various domains, including knowledge work (GDPval: 70.9% wins/ties), science questions (GPQA Diamond: 92.4%), and advanced mathematics (AIME 2025: 100%).

Achieves state-of-the-art performance on rigorous software engineering evaluations, scoring 55.6% on SWE-Bench Pro (four languages) and 80% on SWE-bench Verified (Python). It can reliably debug, implement features, refactor codebases, and handle complex front-end development.

Demonstrates state-of-the-art long-horizon reasoning and understanding of extensive contexts, crucial for complex projects and agentic workflows.

Capable of perceiving images and handling vision-based tasks, including understanding scientific figures, as evidenced by its 88.7% score on CharXiv Reasoning (w/ Python).

Designed to power faster, more reliable agentic workflows, with state-of-the-art long-horizon reasoning and tool-calling performance. It excels at agentic data science and document analysis tasks.

Generates high-quality professional artifacts such as spreadsheets and presentations, often at or above human expert level, with improved sophistication and formatting. It can produce these outputs >11x faster and at <1% the cost of expert professionals.