Discover GPT-5.2
We're introducing GPT-5.2, our most advanced frontier model yet. It is specifically designed for professional knowledge work and empowering long-running autonomous agents to take on complex projects from start to finish.
With significant improvements across the board, GPT-5.2 sets a new standard for AI. It demonstrates extraordinary capabilities in autonomous tool use, interactive programming, and deep contextual reasoning over extensive documents.
Benchmark Performance
The performance increase compared to GPT-5 is massive, particularly in "agentic" disciplines. GPT-5.2 achieves 70.9% on GDPval for agentic reasoning (up from 38.8%) and a staggering 55.6% on SWE-Bench Pro for programming (up from 25.1%). It also pushes boundaries in expert knowledge (92.4% on GPQA Diamond) and higher-education mathematics, scoring 40.3% on FrontierMath compared to its predecessor's 3.2%.
Under the Hood: Core Capabilities
- Agent Performance: Significant improvements in the autonomous use of tools and the execution of complex projects from start to finish.
- Coding: State-of-the-art performance in interactive programming, code reviews, and debugging, confirmed by partners like GitHub/Microsoft and JetBrains.
- Data Analysis: Outstanding in agentic data science tasks and analyzing extensive documents with a deep understanding of long contexts.
- Multimodality: Improved image perception and the ability to seamlessly integrate visual intelligence into extended reasoning tasks.
Sources & Credentials
The information and benchmark statistics regarding GPT-5.2 are sourced directly from the official OpenAI Announcement.