OpenAI Group PBC has unveiled its latest and most advanced large language model, GPT-5.2.
The LLM is available in three variants: Instant, Thinking, and Pro. OpenAI noted that the latter two versions have set new performance records on a range of mathematical benchmarks, and the company claims GPT-5.2 outperforms rival models across multiple domains.
OpenAI evaluated the mid-tier Thinking variant using FrontierMath (Tier 1–3), a benchmark comprising university-level math problems—some of which typically require graduate students several hours to solve. According to OpenAI, GPT-5.2 Thinking correctly solved 40.3% of the problems in the dataset, establishing a new industry record. Additionally, the model achieved a perfect score on the qualifying exam for the International Mathematical Olympiad.
GPT-5.2 Pro, the most powerful version of the LLM, assisted researchers in making novel discoveries in statistical learning theory—a specialized subfield of mathematics. It solved a simplified version of an open problem first posed at a mathematics conference in 2019. OpenAI reported that GPT-5.2 Pro developed this solution entirely without human guidance.
Compared to GPT-5.1, the new model demonstrates significantly improved comprehension of scientific figures and charts. OpenAI assessed this capability using the CharXiv Reasoning benchmark. The Thinking variant correctly interpreted 88.7% of the charts in the dataset—an improvement of over 8 percentage points compared to GPT-5.1 Thinking.
GPT-5.2’s visual reasoning capabilities extend beyond academic contexts. In an internal test, OpenAI provided the model with a low-resolution image of a motherboard, and it successfully identified key components. The model can also analyze business intelligence dashboards, product schematics, and other document types.
OpenAI stated that GPT-5.2 shows marked improvements over its predecessor in front-end development and building visual application components such as user interfaces. The model excels particularly in generating three-dimensional assets and simulations.
The model also delivers enhanced performance on general programming tasks. On SWE-Bench Pro—a suite of complex coding challenges spanning multiple programming languages—GPT-5.2 achieved a record-breaking score of 55.6%. On the Python-only validation subset of SWE-Bench, it scored 80%.
OpenAI has begun rolling out GPT-5.2 to ChatGPT users today and has also made the model accessible to developers via its API.
The entry-level GPT-5.2 is priced at $1.75 per million input tokens and $14 per million output tokens. For applications using the Pro version, these rates increase to $21 and $168 per million tokens, respectively. OpenAI notes that developers can reduce costs for frequently repeated prompts by up to 90% through its caching feature, eliminating the need to regenerate identical responses from scratch.
