OpenAI has launched its most powerful model to date, one the company says “is designed to do far more than answer questions.”
GPT-5.5 is now rolling out to ChatGPT Plus, Pro, Business, and Enterprise users, alongside a more capable variant called GPT-5.5 Pro for Pro, Business, and Enterprise subscribers. OpenAI describes it as a new class of intelligence, built not just to respond, but to work through complex, multi-step tasks from start to finish.
GPT-5.5 is built to chain tasks together, researching, writing, coding, checking its own work, switching between tools, and continuing until a job is complete. OpenAI says the model excels in four key areas: agentic coding, computer use, knowledge work, and scientific research.
What makes this launch technically unusual is that GPT-5.5 delivers these gains without sacrificing speed. Larger, more capable models are typically slower to serve, but OpenAI says GPT-5.5 matches its predecessor’s per-token latency in real-world conditions. The company attributes this to co-designing the model alongside NVIDIA’s GB200 and GB300 NVL72 systems and rebuilding its inference infrastructure from scratch.
For developers, the performance improvements are measurable. On Terminal-Bench 2.0, a benchmark that tests complex command-line workflows, GPT-5.5 scores 82.7%, ahead of Google’s Gemini 3.1 Pro at 68.5% and Anthropic’s Claude Opus 4.7 at 69.4%. On SWE-Bench Pro, which evaluates real-world software issue resolution on GitHub, it reaches 58.6%, solving more problems end-to-end in a single pass than any previous OpenAI model.
Early testers from companies including Cursor, Cognition, and Lovable described a model with stronger conceptual clarity, one that can hold context across large codebases and predict what else in the code might be affected by a change. Dan Shipper, chief executive of Every, said GPT-5.5 was the first coding model he had used with “serious conceptual clarity.” He tasked it with replicating a complex post-launch debugging fix that had originally required one of his best engineers. The model produced the same rewrite. GPT-5.4 could not.
Beyond coding, OpenAI is positioning GPT-5.5 as a tool for everyday professional work. On OSWorld-Verified, a benchmark that tests whether a model can independently operate a real computer environment, it scores 78.7%, on par with Claude Opus 4.7. The company says more than 85% of its own employees use Codex, its agentic platform powered by GPT-5.5, every week across functions including finance, marketing, and data science. In one instance, the finance team used it to review more than 24,000 K-1 tax forms totalling over 71,000 pages, completing the task two weeks faster than the previous year.
Perhaps the most ambitious claims surround scientific research. On BixBench, a benchmark built around real-world bioinformatics tasks, GPT-5.5 achieves 80.5% compared to 74.0% for GPT-5.4. On GeneBench, which tests multi-stage analysis in genetics and quantitative biology, it outperforms its predecessor on tasks that OpenAI says typically take expert researchers multiple days. An internal version of the model also helped discover a new mathematical proof about Ramsey numbers, a longstanding area of research in combinatorics, later verified using the formal proof assistant Lean.
On safety, OpenAI has rated GPT-5.5’s biological and cybersecurity capabilities as “High” under its Preparedness Framework, one level below “Critical.” The company says it is deploying stricter classifiers for high-risk cyber requests and launching a programme called Trusted Access for Cyber, which gives verified security professionals expanded access with fewer restrictions for legitimate defensive work.
GPT-5.5 is available now in ChatGPT and Codex. API access is coming soon, priced at $5 per million input tokens and $30 per million output tokens with a one-million-token context window. GPT-5.5 Pro will cost $30 per million input tokens and $180 per million output tokens. While the model is priced higher than GPT-5.4, OpenAI argues its efficiency gains, including completing the same tasks with fewer tokens, offset the cost difference for most users.
