OpenAI: GPT-4o-mini (batch)

GPT-4o Mini is OpenAI’s small GPT-4o-family model for focused work. It takes text and images and produces text, including Structured Outputs. Its role is useful when an application needs a concise response to a well-specified request: extracting fields, developing a reply from supplied reference material or carrying out a bounded step in a larger process.

For a support response, provide the incoming message, relevant policy excerpts and the facts already established in the conversation. Ask for a short proposed reply and identify any information still needed from the sender. For document extraction, define the fields, explain how to handle missing values and preserve a label that connects each result to its source. The same approach can produce a classification, an edited passage or an explanation of a small code change without treating the model as a substitute for the application’s surrounding logic.

The producer lists a 128,000-token context window, a maximum output of 16,384 tokens and a knowledge cutoff of October 1, 2023. Source material, conversation and generated output occupy that context together. Keep significant earlier facts visible and state the required final artifact. GPT-4o Mini is a separate small model, not an output-length mode of full GPT-4o; its own identity and capabilities should guide the task even where the two models share specifications.

Images can supply a screenshot, a document page, a photograph or a diagram. OpenAI’s vision workflow accepts PNG, JPEG, WebP and non-animated GIF, with low, high or auto detail for this family. Higher detail follows the producer’s tile-based image handling. Keep important labels readable and upright, and explain which part of the image should inform the answer. Fine spatial relationships, exact object counts, unusual image projections and some non-Latin text are documented difficulties. A focused question makes a visual response easier to compare with the original evidence. 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.

The producer’s file workflow passes extracted text and page images from PDFs to vision-capable models. Other office and text files contribute extracted text, so their embedded images and charts need a PDF representation when those visuals matter. Spreadsheet augmentation uses up to the first 1,000 rows per sheet. Each file must be under 50 MB, with a combined request allowance of 50 MB. Chat Completions accepts PDF file parts; the documented Responses workflow accepts a broader set of file formats. Preserve page, row or section identifiers in the material you use for extraction.

Streaming returns text incrementally. Structured Outputs constrains a completed response to the supported JSON Schema subset, whereas a request for JSON mode only ensures valid JSON. Name the required fields and permitted values, including an explicit representation for missing evidence. Applications need to handle refusals and incomplete responses separately. A schema can make the artifact suitable for the next stage, but the values still need to reflect the actual source rather than an invented completion of an empty field.

GPT-4o Mini supports function calling and the producer’s tool-oriented workflows with web search, file search, code interpretation and MCP connections. Describe an available function, let the application execute the requested operation and return its observation before the model develops a follow-up answer. OpenAI also supports a separate image-generation tool in the configured Responses workflow. Images produced by that tool are separate artifacts; the model’s native documented output remains text. Search and tool use are configured activities, rather than evidence that every reply consulted current information.

Predicted Outputs supports regenerating a largely known document or code file with a small change. Provide the previous content and the precise edit, then compare the resulting artifact with that request. The prediction path supports text only and does not combine with function calling; logprobs, max_completion_tokens and several other controls are also excluded. An ordinary tool-using request is a different path when the reply depends on a fresh lookup or calculation.

OpenAI documents fine-tuning for this model and describes using a larger model’s outputs to teach a smaller GPT-4o Mini model a specialised task. This distillation workflow produces a customised model; it does not change the identity of a normal Mini response. For an ordinary request, establish the task and output format with clear examples, then use a focused follow-up to revise the same result. A useful final artifact remains connected to the source facts, the intended reader and the application step that consumes it.