OpenAI rolls out GPT-5.5 with coding & research gains

OpenAI has released GPT-5.5 for ChatGPT and Codex users, with API access to follow.

The rollout covers Plus, Pro, Business and Enterprise tiers in ChatGPT and Codex. GPT-5.5 Pro is also being introduced for Pro, Business and Enterprise users in ChatGPT. In Codex, GPT-5.5 is available on Plus, Pro, Business, Enterprise, Edu and Go plans, with a 400,000-token context window.

OpenAI says the model is built to handle broader, multi-step tasks with less direct prompting. According to the company, GPT-5.5 performs strongly in coding, online research, data analysis, document creation, spreadsheet work and software operation, while maintaining the same per-token latency as GPT-5.4 in real-world serving.

It also uses fewer tokens than GPT-5.4 on the same Codex tasks, which OpenAI says makes it both more efficient and more effective for coding and knowledge work.

Coding focus

A large part of the update centres on software engineering. OpenAI says GPT-5.5 scored 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and outperformed GPT-5.4 on its internal Expert-SWE evaluation for long-horizon coding tasks.

These tests are designed to measure how models handle command-line workflows, resolve GitHub issues and manage coding work that can take humans many hours. OpenAI says GPT-5.5 improved on GPT-5.4 across all three evaluations while using fewer tokens.

The company says those gains are visible in Codex, where the model can handle implementation, refactoring, debugging, testing and validation. Early testing also suggested GPT-5.5 was better at retaining context across larger systems, reasoning through ambiguous failures and carrying changes through a wider codebase.

Work tasks

OpenAI also presents GPT-5.5 as a tool for broader office work. It says the same improvements that help with coding also support research, information gathering, output checking and turning raw material into documents, spreadsheets and slide presentations.

Within its own operations, more than 85% of staff use Codex every week across software engineering, finance, communications, marketing, data science and product management, according to OpenAI. The company cited examples including analysis of six months of speaking request data by its communications team, a review of 24,771 K-1 tax forms totalling 71,637 pages in finance, and automated weekly business reporting that it says saved one employee five to 10 hours a week.

For benchmarked knowledge work, OpenAI says GPT-5.5 scored 84.9% on GDPval, 78.7% on OSWorld-Verified and 98.0% on Tau2-bench Telecom without prompt tuning. It also reported scores of 60.0% on FinanceAgent, 88.5% on internal investment banking modelling tasks and 54.1% on OfficeQA Pro.

Research claims

OpenAI says GPT-5.5 also posted gains in scientific and technical research workflows, including genetics, quantitative biology and bioinformatics. It cited improvements over GPT-5.4 on GeneBench and said GPT-5.5 achieved leading performance among models with published scores on BixBench.

The company also said an internal version of GPT-5.5, used with a custom harness, helped discover a new proof related to off-diagonal Ramsey numbers, a topic in combinatorics. OpenAI says the result was later verified in Lean.

Infrastructure changes

To support the model at GPT-5.4 latency, OpenAI says it redesigned parts of its inference system and co-designed, trained and served GPT-5.5 on NVIDIA GB200 and GB300 NVL72 systems. It also says Codex analysed weeks of production traffic patterns and wrote custom heuristic algorithms for load balancing and partitioning, increasing token generation speeds by more than 20%.

Safety measures

OpenAI says GPT-5.5 is being released with what it describes as its strongest safeguards so far. The company says it evaluated the model under its safety and preparedness frameworks, carried out targeted testing for advanced cybersecurity and biology risks, worked with internal and external red teamers, and gathered feedback from nearly 200 trusted early-access partners.

According to OpenAI, GPT-5.5 is rated High for biological, chemical and cybersecurity risk under its Preparedness Framework. The company added that the model did not reach what it classifies as a Critical cybersecurity capability level, but said it marks a step up from GPT-5.4 and is launching with tighter controls around higher-risk cyber activity and repeated misuse.

For developers, OpenAI says gpt-5.5 will be offered through the Responses and Chat Completions APIs at USD $5 per one million input tokens and USD $30 per one million output tokens, with a one million-token context window. It says gpt-5.5-pro will also be released through the API at USD $30 per one million input tokens and USD $180 per one million output tokens.

OpenAI says GPT-5.5 is priced above GPT-5.4, but adds that it has tuned Codex so the newer model generally delivers better results while using fewer tokens.