OpenAI: GPT-4o (batch)

GPT-4o is OpenAI’s versatile GPT model for text and image understanding, writing and application work. The “o” stands for “omni”, while this documented model accepts text and images and returns text, including Structured Outputs. A practical request can combine a written brief with photographs, screenshots or document pages and ask for a clear explanation, an edited artifact or defined fields for another application stage.

For a report, provide the source documents and the decision the report should support. Identify the relevant dates, units and evidence labels, then request a comparison that separates observations, assumptions and missing information. For a visual task, point to the object, chart or interface state that matters and include the surrounding context. For a code change, supply the affected code, the desired behavior and the constraints that must remain intact. These are concrete ways to use the model’s text and image capabilities without relying on the prompt to imply the whole task.

OpenAI’s current GPT-4o profile lists a 128,000-token context window, a maximum output of 16,384 tokens and a knowledge cutoff of October 1, 2023. The producer lists several dated GPT-4o snapshots separately. The conversation, input material and generated answer share the context budget, so a large document collection still needs organisation and room for its final artifact. Facts that change or come from a private system should be supplied directly or obtained through an explicitly configured tool.

The vision guide accepts PNG, JPEG, WebP and non-animated GIF. GPT-4o supports low, high and auto image detail; high and auto follow tile-based processing. Enlarge small labels, keep text upright and preserve the legend or surrounding diagram when interpreting a chart. The documented limitations include precise spatial localisation, approximate object counts, some non-Latin writing and panoramic or fisheye views. The model can also describe an image incorrectly, so retain the source and ask for the specific observations that support an interpretation. In OpenAI’s documented vision workflow, the total request payload is limited to 512 MB and a request can include up to 1,500 images; the images and text must also fit the model’s context.

OpenAI’s file-input workflow gives a vision-capable model both extracted text and page images from PDFs. Non-PDF office documents supply extracted text rather than their embedded charts, and spreadsheets are augmented from up to the first 1,000 rows per sheet. Each file must be under 50 MB, with 50 MB combined across a request. Chat Completions supports PDF file parts; the documented Responses workflow accepts additional file formats. Choose a PDF representation where visual fidelity matters and keep source identifiers visible in the result.

Streaming can return the response as it is generated. Structured Outputs constrains a completed response to the supported subset of JSON Schema, making it useful for document extraction or an application object with defined fields. JSON mode only addresses JSON validity. Define the expected types, allowed values and treatment of absent evidence, then handle explicit refusals or incomplete responses separately. A schema does not establish whether a value is factually correct, so the requested fields should remain traceable to the supplied material.

The producer supports function calling, web search, file search, code interpretation and MCP connections. Custom functions need meaningful descriptions and argument schemas; the application executes the requested action and returns its result for the next model step. GPT-4o can also direct a separate image-generation tool in a configured Responses workflow. A tool-created image is its own artifact, while GPT-4o’s native documented response is text. An agent’s final explanation should use actual retrieved or calculated observations rather than treating a proposed tool call as a completed operation.

Predicted Outputs is available for GPT-4o when regenerating largely known text or code with small changes. Supply the existing content and a precise modification, then inspect the returned artifact. This text-only path does not combine with function calling and excludes logprobs, max_completion_tokens and several other controls. A tool-based request is a separate choice when the answer depends on new information from an external operation.

OpenAI documents fine-tuning for GPT-4o, and its optimisation guide specifically names the August 6, 2024 snapshot for vision fine-tuning with image examples. That snapshot-specific customisation workflow is distinct from an ordinary request to the model alias. For daily writing, analysis or extraction, begin with a useful brief and an example of the desired output. A follow-up can resolve one missing fact, revise a field or refine the same artifact while preserving the task’s evidence and constraints.