OpenAI has officially released GPT-5.5, a model engineered to move beyond conversational responses toward autonomous task execution. Described as the “next step” in agentic AI, the model is designed to navigate complex, multi-part workflows—such as debugging code, operating software, and conducting multi-step research—with minimal human oversight. The model is now rolling out to Plus, Team, and Enterprise users across ChatGPT and the Codex environment, where it is already being used to automate engineering tasks and long-horizon knowledge work.
The model demonstrates significant gains in “computer use” and coding, achieving 82.7% accuracy on Terminal-Bench 2.0, which evaluates complex command-line planning. While it trails slightly behind Claude Opus 4.7 on GitHub issue resolution (SWE-Bench Pro), OpenAI notes that GPT-5.5 is more token-efficient and maintains the same per-token latency as its predecessor, GPT-5.4. Internal deployments at NVIDIA have already reportedly cut debugging cycles from days to hours, utilizing the model to ship end-to-end features directly from natural language prompts.
To support the massive compute density required for these agentic workflows, GPT-5.5 was co-designed and served on NVIDIA’s latest GB200 and GB300 NVL72 Blackwell systems. This hardware integration allows the model to process tasks more efficiently, using significantly fewer tokens than previous iterations to achieve higher-quality outcomes. OpenAI has also implemented its strictest safety classifiers to date, specifically targeting cybersecurity and biosecurity risks, as the model’s ability to interact with real-world computer systems increases.