Introducing GPT-5.2

OpenAI has unveiled GPT-5.2, a state-of-the-art model optimized for expert knowledge work and long-running agents. This model is designed to maximize economic value by significantly enhancing capabilities in spreadsheet creation, presentation building, code writing, image recognition, long-context understanding, and executing complex multi-step projects.

GPT-5.2 Thinking achieved new SOTA (State-of-the-Art) results across various benchmarks. Notably, in the GDPval test covering 44 occupations, it demonstrated performance on par with or exceeding human experts (70.9% win or tie), proving it can perform tasks at 11x the speed of experts and at under 1% of the cost.

Key performance metrics are as follows:

• GDPval (knowledge work): 70.9%
• AIME 2025 (competition math): 100.0%
• ARC-AGI-2 (abstract reasoning): 52.9%
• SWE-Bench Pro (software engineering): 55.6%

Notion, Box, and Zoom have confirmed GPT-5.2's outstanding reasoning and tool-use capabilities, while Cognition and JetBrains reported dramatic improvements in agentic coding performance.

For ChatGPT users (paid plans), GPT-5.2 Instant, Thinking, and Pro models will be rolled out sequentially, and the API is now immediately available to all developers.