openai / gpt-5.4

## Overview GPT-5.4 by OpenAI is a long-context reasoning model that reads text and optional images, then writes a text response. It supports both single-turn prompts and multi-turn chat messages. You can dial up reasoning depth when you need multi-step analysis. This helps you turn screenshots, charts, and documents into clear explanations or extracted details. ## What you can build - Visual Q&A over screenshots, photos, and UI mockups - Chart and table reading, including extracting key numbers into clean text - Document review from images, like scanned pages or slide screenshots - Image-based troubleshooting for error dialogs, logs, and configuration screens - Code generation and refactors, guided by a spec and optional screenshots - Step-by-step workflows that require longer context and careful tracking ## Inputs - A required instruction written in plain text that explains the task and the expected format. - Optional chat history as a JSON-formatted list of messages (for example: user and assistant turns). When you provide message history, the single-turn prompt and system instructions are ignored. - Optional system-level instructions in plain text that set rules for tone, formatting, and boundaries. - Optional one or more image files for analysis. Supported formats include PNG, JPEG/JPG, WEBP, and non-animated GIF. - Optional image batching: you can include many images in one request, up to 500 images total, with a total payload up to 50 MB. - Optional end-user and session identifiers (string or number) to support chat history and session continuity. - Optional reasoning depth selection from fast to maximum-depth. Higher reasoning depth can use more internal reasoning tokens and may need a higher output token budget. - Optional verbosity control to keep responses short, balanced, or detailed. - Optional maximum output length, expressed as a token budget for the generated response. - Optional web search toggle to allow the model to browse the web when your prompt calls for it. ## Outputs The model returns generated text as an ordered list of text chunks. Join the chunks to form the final response. The text can include structured formatting like headings, bullet lists, tables in Markdown, or code blocks, depending on your instructions. ## Recommended settings - Use fast or low reasoning depth for short tasks like captioning a screenshot. - Use high or maximum reasoning depth for multi-step tasks like reconciling values across charts. - If you choose high reasoning depth, increase the output token budget to avoid empty or cut-off replies. - Set low verbosity for concise extraction and high verbosity for explanations and step-by-step plans. ## Limitations - GPT-5.4 accepts image input but does not support audio or video inputs. - The maximum context window is 1,050,000 tokens, and the maximum output is 128,000 tokens. - The model knowledge cutoff is August 31, 2025, so it may miss newer facts without web search. - Vision quality depends on the input image. Blurry screenshots, tiny text, glare, or heavy compression can cause missed details. - Images with watermarks or logos can be rejected. Images must also be clear enough for a human to understand. - Images with text in non-Latin alphabets may perform worse than Latin-script text. ## Safety & compliance OpenAI trains GPT-5.4 Thinking with policies that target disallowed content categories like harassment, sexual content, hate, self-harm, and wrongdoing. For computer-use scenarios, OpenAI uses a confirmation policy approach that can require extra confirmation for higher-risk actions. OpenAI also applies added cyber safeguards, including classifier-based blocking and account-level monitoring for high-risk cyber misuse.