gpt-5.5

Overview. GPT-5.5 is an instruct model in OpenAI's GPT-5 family, released in April 2026, with a 1,050,000-token context window and a December 2025 training cutoff. It accepts text, image, and file input and returns text output, making it a fit for teams that need to reason over very long documents or sessions within a single call. A developer would reach for it when the family's smaller mini/nano-tier models (e.g., gpt-5.4-mini) fall short on context length but a dedicated reasoning-tagged variant isn't required.

Key strengths. OpenAI's Artificial Analysis benchmarking (as surfaced on the model's endpoint) puts GPT-5.5's coding index at 74.9 and intelligence index at 56.3, with an agentic index of 47.4 — solid coding performance paired with reasonable general reasoning and tool-use capability. It supports a 'reasoning' effort control (none through xhigh) in place of temperature, letting callers tune reasoning depth per request. Note that within its own family, GPT-5.6-luna and GPT-5.6-sol now score higher on quality index, so GPT-5.5 is a strong mid-to-upper-tier choice rather than the family's top performer.

When to use it. Use GPT-5.5 for complex multi-step reasoning, long-document analysis, and tool orchestration where the 1M+ token context window matters — contract review, large codebase question-answering, or multi-file research synthesis. It's built for interactive, API-driven usage with standard chat-completions semantics, including system prompts and tool/function calling. If your task requires the family's highest reasoning score, consider GPT-5.6-sol instead.

Limitations. Maximum output is capped at 128,000 tokens even though the input context window is over 1,000,000 tokens, so very long generations still need to be chunked. GPT-5.5 has no audio input or output, and its training data ends in December 2025. It's also no longer the top-scoring instruct variant in its own family, having been superseded on quality index by newer siblings.

Sovereignty note. GPT-5.5 is built by OpenAI, a US-based creator with confirmed CLOUD Act exposure, meaning data processed through this model carries US legal reach regardless of where Inference is routed. It is accessed only via API — there is no self-hosted or Sovereign deployment path for this model.