OpenAI has released GPT-5.5, its most advanced AI model yet, designed to handle complex, multi-step tasks with greater autonomy and efficiency. The model represents a significant leap in AI capability, particularly in areas requiring sustained reasoning and real-world computer interaction.
GPT-5.5 excels at understanding messy, multi-part tasks and can plan, use tools, check its work, and navigate through ambiguity without constant supervision. The model shows particular strength in coding, scientific research, and knowledge work where progress depends on reasoning across context and taking action over time.
Major improvements in coding and software development
GPT-5.5 delivers state-of-the-art performance in programming tasks. On Terminal-Bench 2.0, which tests complex command-line workflows requiring planning and tool coordination, it achieves 82.7% accuracy. The model also reaches 58.6% on SWE-Bench Pro, which evaluates real-world GitHub issue resolution.
The coding improvements extend beyond benchmarks. Early testers report that GPT-5.5 shows stronger conceptual clarity and can understand system architecture at a deeper level. Dan Shipper, CEO of Every, described it as “the first coding model I’ve used that has serious conceptual clarity.” In one test case, GPT-5.5 successfully rewound and fixed a complex post-launch bug that had required days of debugging and a complete system rewrite by human engineers.
Key coding capabilities include:
- Better context retention across large codebases
- Improved debugging and error analysis
- More accurate understanding of system dependencies
- Reduced token usage while achieving higher quality outputs
Scientific research and mathematical breakthroughs
The model shows significant gains in scientific workflows that require exploring ideas, gathering evidence, and iterating through multi-step research processes. GPT-5.5 improved substantially on GeneBench, which focuses on multi-stage scientific data analysis in genetics and quantitative biology.
In a notable achievement, an internal version of GPT-5.5 helped discover a new proof about Ramsey numbers, one of the central objects in combinatorics. The model found a proof of a longstanding asymptotic fact about off-diagonal Ramsey numbers, which was later verified in Lean theorem prover.
Researchers are using GPT-5.5 as a genuine research partner. Derya Unutmaz, an immunology professor at Jackson Laboratory, used the model to analyze a gene-expression dataset with 62 samples and nearly 28,000 genes, producing detailed research findings that would have taken his team months to complete.
Enhanced computer use and knowledge work
GPT-5.5’s ability to understand intent translates into more natural computer interaction. The model can move through the full loop of knowledge work: finding information, understanding what matters, using tools, checking outputs, and turning raw material into finished products.
At OpenAI, more than 85% of the company now uses the model weekly across various functions including engineering, finance, marketing, and product management. Teams report significant time savings, with one employee automating weekly business reports and saving 5-10 hours per week.
The model performs well on knowledge work benchmarks:
- 84.9% on GDPval, which tests knowledge work across 44 occupations
- 78.7% on OSWorld-Verified, measuring computer environment operation
- 98.0% on Tau2-bench Telecom for complex customer service workflows
Cybersecurity capabilities and safety measures
GPT-5.5 includes significantly enhanced cybersecurity capabilities, which OpenAI is managing through what it calls “industry-leading safeguards.” The company has implemented stricter controls around higher-risk activities and sensitive cyber requests.
To balance capability with safety, OpenAI is expanding access through several programs:
- Trusted Access for Cyber program for verified users
- Special access for organizations defending critical infrastructure
- Collaboration with government partners on public security systems
The model achieved 81.8% on CyberGym benchmarks and 88.1% on internal Capture-the-Flag challenges, representing a clear step up from previous versions.
Availability and performance improvements
Despite being more capable, GPT-5.5 matches GPT-5.4’s response speed while using fewer tokens to complete tasks. This efficiency improvement came from co-designing the model with NVIDIA GB200 and GB300 systems and using AI assistance to optimize the serving infrastructure itself.
The model is rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex today. API access will follow soon, with pricing at $5 per 1M input tokens and $30 per 1M output tokens. A more advanced GPT-5.5 Pro version is available for Pro, Business, and Enterprise users.
GPT-5.5 represents a significant step toward AI systems that can genuinely collaborate on complex, real-world tasks rather than simply answering questions. The combination of improved reasoning, autonomous operation, and maintained speed suggests we’re entering a new phase of AI capability where models can serve as true working partners across technical and creative domains.
