OpenAI Group PBC today launched GPT-5.2, its newest and most capable large language model.

The LLM is available in three versions: Instant, Thinking and Pro. OpenAI says that the latter two editions provide record-setting performance across many mathematical tasks. The company claims that GPT-5.2 also outperforms rivals in other areas. 

OpenAI tested the mid-range Thinking version of the model using FrontierMath (Tier 1-3), a benchmark dataset that comprises college-level math problems. Some of the questions take graduate students several hours to solve. OpenAI says that GPT-5.2 Thinking solved 40.3% of the problems in the dataset correctly, a new industry record. Additionally, the model achieved a perfect score on a qualifying exam for the International Mathematical Olympiad.

GPT-5.2 Pro, the LLM’s most capable version, helped researchers make a new discovery in a mathematical subfield called statistical learning theory. It solved a simple version of an open problem that was floated during a 2019 math conference. According to OpenAI, GPT-5.2 Pro developed the answer without pointers from humans on how it should go about the task.

Compared with GPT-5.1, the model is better at understanding charts in scientific papers. OpenAI evaluated GPT-5.2’s performance in that area using a benchmark called CharXiv Reasoning. The Thinking version of the model correctly interpreted 88.7% of the charts in the benchmark dataset, a more than 8% improvement over GPT-5.1 Thinking.

GPT-5.2’s visual reasoning features also lend themselves to other tasks. In one internal test, OpenAI staffers provided the model with a low-resolution image of a motherboard and successfully used it to identify key components. GPT-5.2 can also analyze business intelligence dashboards, product diagrams and other files. 

OpenAI says that the model is significantly better than its predecessor at front-end development, or the task of building visual application components such as interfaces. GPT-5.2 is particularly adept at creating three-dimensional assets such as simulations. 

The model also brings performance improvements across other programming tasks. OpenAI says that GPT-5.2 achieved a record 55.6% score on SWE-Bench Pro, a collection of difficult coding tasks spanning multiple programming languages. It scored 80% on the Python-only SWE-bench Verified version of the benchmark.

OpenAI started rolling out GPT-5.2 to ChatGPT today. It also made the LLM available through its application programming interface for developers. 

The entry-level GPT-5.2 model is pricing at $1.75 per million input tokens and $14 per million output tokens. Those rates jump to $21 and $168, respectively, for applications that use the Pro version of the LLM. OpenAI says that developers can reduce output costs by up to 90% using a caching feature that saves frequent prompt answers, which removes the need to generate from scratch in response to every request.

Photo: Focal Foto/Flickr