OpenAI releases GPT-5.5, a ground-up retrained model with native omnimodal processing

OpenAI has released GPT-5.5, the first fully retrained base model since GPT-4.5, featuring a unified architecture for text, images, audio, and video. The model was co-designed with NVIDIA's GB200 and GB300 NVL72 systems to maintain low latency despite increased capability.

• GPT-5.5 achieves state-of-the-art results on Terminal-Bench 2.0 (82.7%) and FrontierMath (51.7%), leading Claude Opus 4.7 in these categories.
• It demonstrates a significant 37-point improvement in long-context performance, reaching 74.0% on MRCR v2 at 512K–1M token contexts.
• The model improves abstract reasoning scores to 85.0% on ARC-AGI-2 and cybersecurity capabilities to 93.33% pass rate on cyber range evaluations.
• Claude Opus 4.7 retains leads in SWE-bench Pro (64.3% vs 58.6%) and tool orchestration benchmarks like MCP Atlas.

The release marks a shift toward agentic coding and complex workflow automation, with OpenAI reporting that internal infrastructure changes increased token generation speeds by over 20%.