GPT-5.5

GPT-5.5 is OpenAI's first fully retrained base model since GPT-4.5, internally codenamed "Spud," and represents a ground-up rebuild rather than another incremental update on the prior architectural foundation. According to third-party analysis, it moves away from stitched-together modality handlers and instead processes text, images, audio, and video end-to-end within a single unified architecture. That omnimodal redesign is the headline structural change of the release, intended to let one model reason across formats without passing inputs between specialized submodels.

Beyond architecture, the release is framed around tighter hardware and infrastructure integration. The model was reportedly co-designed alongside NVIDIA's GB200-class rack-scale systems, which third-party reporting credits with helping GPT-5.5 match the per-token latency of its predecessor despite a clear capability jump. OpenAI's own Codex tooling was also used to rewrite portions of the serving stack before launch, reportedly yielding faster token generation once the model went live. Taken together, the picture is a flagship aimed at long-context, multi-format assistant workloads, where unified multimodal understanding, sustained reasoning, and high-throughput generation matter more than any single benchmark number.