gpt-5.4

Overview. GPT-5.4, marketed by OpenAI as 'GPT-5.4 Thinking', is a proprietary chain-of-thought reasoning model released March 5, 2026. It offers a 1,050,000-token context window with up to 128,000 tokens of output, and accepts text, image, and file input natively. Developers reach for it when a task needs sustained reasoning over very large documents or codebases rather than a quick conversational reply.

Key strengths. Independent benchmark composites place GPT-5.4 near the top of its own model family for coding ability, with a coding index around 71 out of 100, and a solid general intelligence index in the low 50s. Its reasoning effort is configurable across five levels, from none to maximum ('xhigh'), letting a developer trade latency for depth of Inference on a per-request basis. The 1M-token context window makes it well suited to whole-repository or whole-document analysis rather than short prompts.

When to use it. Use GPT-5.4 for complex coding, multi-step reasoning, or large-document analysis where the input can run into hundreds of thousands of tokens and a slower, more deliberate answer is acceptable. Its agentic-task score is measurably lower than its coding score, so it fits better as a reasoning engine behind a workflow than as the sole driver of open-ended multi-tool agent loops. The 'reasoning' complexity fit and configurable effort levels make it a good match for review, analysis, and planning tasks rather than fast interactive chat.

Limitations. GPT-5.4 behaves as a batch-latency model: extended internal reasoning traces mean it is not tuned for snappy, low-latency interactions. It has no audio input or output path, and OpenAI has not disclosed parameter count, architecture, or a firm training cutoff beyond an approximate mid-2025 date. It is closed-weight, so it cannot be self-hosted or inspected, and it was retired from OpenAI's consumer ChatGPT product in August 2026, though it remains available through the API that this Platform routes to.

Sovereignty note. GPT-5.4 is built and operated by OpenAI, a US company, so it does not carry Sovereign, EU-residency guarantees and is subject to potential CLOUD Act exposure. Usage is governed by OpenAI's published usage policies rather than a per-model licence.