OpenAI Group PBC today launched a new large language model that is significantly better than its predecessors at solving math problems and writing code.

GPT-5.5 is rolling out a week after rival Anthropic PBC released its latest LLM. OpenAI is offering the model in two flavors: a standard version and a more capable, significantly pricier edition called GPT-5.5 Pro.

The company says both variants deliver output quality improvements in multiple areas. The standard edition of GPT-5.5 is more adept than its predecessor at computer use tasks and knowledge work. GPT-5.5 Pro, in turn, provides particularly large quality gains across business, legal, education and data science use cases.

GPT-5.5 is also better at interpreting ambiguous instructions. Historically, LLM users had to describe each step of the task they sought to automate or risk output errors. By contrast, GPT-5.5 can automatically figure out details such as how to use an MCP server even if the user doesn’t provide an explanation.

OpenAI compared GPT-5.5 with Claude Opus 4.7, the new LLM that Anthropic debuted last week, across more than a dozen benchmarks. The standard and Pro editions of the former model performed better across many of the tests.

One of the most difficult benchmarks in OpenAI’s test suite is FrontierMath Tier 4. It comprises dozens of postdoctoral-level math problems that can take a human expert upwards of days to solve. GPT 5.5 Pro scored 39.6%, nearly double the 22.9% that Claude Opus 4.7 achieved.

OpenAI says that a customized version of GPT-5.5 helped researchers discover a new proof, a series of equations that confirms a mathematical theorem. The proof related to objects known as Ramsey numbers. Such objects are a major focus of a mathematical field called combinatorics that has broad computer science applications.

According to OpenAI, GPT-5.5 is also better than competing models at many programming tasks. The standard version of the LLM achieved a 82.7% score on Terminal-Bench 2.0, which measures LLMs’ ability to use command line tools. Claude Opus 4.7 scored 69.4%.

OpenAI says it has already put GPT-5.5’s coding skills to use internally. The LLM helped optimize the software that manages the infrastructure on which it runs. That hardware comprises Nvidia Corp.’s GB200 and GB300 NVL72 systems, which include the chipmaker’s Blackwell B200 and Blackwell Ultra graphics processing units, respectively. 

GPUs have significantly more cores than a central processing unit. OpenAI’s infrastructure management software groups the LLM requests that are sent to a GPU into batches, or chunks, and distributes them across the chip’s cores. According to the company, GPT-5.5 developed a more efficient way of going about the task that increased token generation speeds by over 20%.

The model is also proficient at less technical tasks. It set a record on GDPval, a benchmark dataset that tests LLMs’ ability to complete economically valuable tasks across 44 fields. Notably, the standard version of GPT-5.5 bested both the Pro edition and Claude Opus 4.7 with a 84.9% score. 

GPT 5.5 is available in ChatGPT and Codex for users with Plus, Pro, Business, and Enterprise subscriptions. GPT-5.5 Pro is only available in the latter 3 plans through ChatGPT. OpenAI will bring the LLM to its application programming interface “very soon.”

Image: OpenAI