OpenAI launches GPT-5.5 for agentic coding and computer-based work

OpenAI launched GPT-5.5 on April 23, positioning the model for agentic coding, computer use, knowledge work, and scientific research. It began rolling out in ChatGPT and Codex, with GPT-5.5 Pro available on selected paid plans. OpenAI updated the post on April 24 when GPT-5.5 and GPT-5.5 Pro reached the API.

The release emphasizes work across tools

OpenAI’s examples focus on tasks that require planning, browsing, coding, document creation, data analysis, and checking results over time. In Codex, the model launched with a 400K context window and a faster serving option. The company reported improved scores over GPT-5.4 on several coding and computer-use evaluations while using fewer tokens on Codex tasks. Those benchmark results should be read with their harnesses, model settings, and the company’s noted memorization concern for SWE-Bench Pro.

The model also helped optimize parts of its serving stack. OpenAI says Codex analyzed production traffic and wrote partitioning heuristics, while GPT-5.5 contributed to inference improvements. The release classified its biological and cybersecurity capabilities as High under the Preparedness Framework and introduced tighter cyber controls, repeated-misuse protections, and a trusted-access path for verified defenders.

Migration requires task-level evaluation

Developers should avoid treating a model upgrade as a drop-in quality guarantee. Replay representative tool calls, long-context tasks, permission failures, structured outputs, and cancellation paths. Measure completed-task cost rather than token price alone, and inspect whether shorter reasoning changes auditability or error recovery.

The API arrived a day after the initial product release, so availability must be tied to the dated deployment surface. Teams should pin model identifiers where reproducibility matters and keep rollback criteria ready while prompts, tools, and evaluators are recalibrated.