GPT-5.5 Pro (xhigh) is a specialized configuration of OpenAI's frontier language model family, released on April 23, 2026. Codenamed Spud, it represents the first complete retraining of the underlying base model architecture since the 4.5 series, distinguishing it from the incremental post-training updates of the earlier GPT-5.x versions. This model is designed for high-accuracy reasoning, complex retrieval, and agentic workflows, utilizing a native omnimodal architecture that processes text, images, audio, and video within a single unified system.
The xhigh designation refers to the maximum reasoning effort level available for the model, specifically optimized for the most demanding mathematical proofs, research, and coding tasks. In performance benchmarks, this configuration achieved 82.7% on Terminal-Bench 2.0 and 73.1% on the internal Expert-SWE benchmark, which evaluates long-horizon coding tasks. The model is noted for its ability to handle complex problem-solving with significantly higher efficiency, often requiring fewer output tokens than its predecessors to complete identical tasks.
Architecturally, GPT-5.5 Pro was co-designed with NVIDIA's rack-scale infrastructure, including the GB200 and GB300 systems. This integration allows the model to match the per-token latency of earlier models despite its significantly larger capability set. It also incorporates self-improving serving infrastructure, where model-generated heuristics were used to optimize load-balancing and inference speeds. The model provides a 1 million token context window in the API, facilitating the processing of entire codebases or massive document sets in a single prompt.
Official prompt guidance for the GPT-5.5 series recommends a more outcome-oriented interaction style. Users are advised to start with concise prompts that define the target outcome and verification criteria, allowing the model's enhanced steerability to determine the optimal execution path. This approach reduces the need for the extensive step-by-step process scaffolding required by previous generations, as the model is more interactive and responsive to corrections during agentic loops.