GPT-5.5 (xhigh)

GPT-5.5 (xhigh) is a frontier large language model developed by OpenAI and released on April 23, 2026. It represents the highest-tier reasoning configuration within the GPT-5.5 series, specifically engineered for high-stakes analytical tasks, scientific research, and complex problem-solving. Positioned as a significant retraining of the company's core architecture, this variant is optimized for "agentic" workflows, allowing the model to plan, verify, and execute long-horizon tasks autonomously for extended periods.

Performance evaluations indicate that the model achieves top-tier results across several specialized benchmarks, including a 93.5% score on GPQA (Graduate-Level Google-Proof Q&A) and a leading score on the Terminal-Bench Hard evaluation. The "xhigh" setting enables maximum internal processing effort, which is particularly effective for code generation, where it maintains a high coding index by performing internal debugging and verification before generating a final response. This capability makes it suitable for advanced software engineering and mathematical applications.

The model is natively multimodal, accepting both text and image inputs while providing high-fidelity text outputs. It supports an expansive context window of 1,000,000 tokens, with a specialized output limit of 128,000 tokens. To handle the increased scale of the model without sacrificing speed, OpenAI co-designed the inference stack for NVIDIA GB200 and GB300 systems, achieving per-token latency parity with previous, less capable versions.

GPT-5.5 (xhigh) introduces granular controls for reasoning effort, allowing users to prioritize depth of thought over immediate output. In this configuration, the model often provides a brief overview of its reasoning approach before executing a task, permitting users to interject and redirect the process if necessary. This interactive reasoning cycle is intended to facilitate a more collaborative relationship between the user and the system during complex document creation or research synthesis.