OpenAI positions GPT-5.5 as a ground-up retrain rather than a point update on the previous base, with a fresh architecture and new training distribution that third-party reviewers describe as the first fully rebuilt foundation since GPT-4.5. The release was announced on 23 April 2026 under the internal codename "spud," and the model is offered in multiple tiers, including a default surface alongside Thinking and Pro variants. API access exposes configurable reasoning effort levels ranging from non-reasoning up through medium, high, and xhigh, letting teams dial deliberation depth for each workload. The multimodal stack carries over the text and image inputs familiar from the wider GPT-5 family, while outputs stay text-based for predictable integration into existing pipelines.
Headline benchmark performance centers on a Terminal-Bench 2.0 score of 82.7%, paired with an ARC-AGI-2 result of 85.0%, marking a meaningful step forward for agentic coding and multi-step tool use. A one-million-token that quick-info value window places the model alongside the longest-context systems in the field, enabling whole-codebase reasoning and long-document analysis without aggressive chunking. In practice, the combination of tiered reasoning modes and the extended that quick-info value makes GPT-5.5 a strong fit for complex software engineering workflows, multi-document research assistants, and agent loops that need both breadth of memory and careful step-by-step planning. Teams building production agents or long-context retrieval pipelines get a flagship option designed explicitly for that class of work.
