OpenAI: GPT-4.1 Nano (batch)

GPT-4.1 Nano is OpenAI’s fastest member of the GPT-4.1 family, designed for instruction following and tool calling without a separate reasoning step. It accepts text and images and produces text. Nano suits a well-defined operation where the application needs a compact answer, a set of fields or the next tool request rather than an expansive investigation.

A useful example is sorting incoming messages against a supplied set of categories. Explain each category, show how ambiguous or out-of-scope messages should be represented, and ask for the category together with the source phrase that supports it. For extraction, define field names, permitted values and the treatment of missing information. For a short writing task, supply the facts and ask for the specific passage that the next application stage needs. The examples and constraints should be small enough to remain clear even when the source collection is large.

The producer lists a 1,047,576-token context window, a maximum output of 32,768 tokens and a knowledge cutoff of June 1, 2024. These specifications match the window of GPT-4.1 Mini while preserving Nano’s separate model identity and responsiveness-oriented role. Label documents and conversation sections, include only the context relevant to the operation, and reserve space for the complete answer. Retrieval results or supplied reference text can provide facts that are newer than the model’s learned knowledge.

Nano supports streaming and Structured Outputs. A schema-constrained final response serves a different purpose from a function call: one gives the application a completed artifact, while the other asks it to perform an operation. Structured Outputs supports a subset of JSON Schema; define required properties and a clear representation for unknown values. Handle refusals and incomplete responses separately, and check the extracted values against the input rather than equating a valid shape with a correct conclusion.

The documented tool set includes function calling, file search, code interpretation, MCP connections and a separate image-generation tool. The application supplies tool descriptions and schemas, executes custom functions and returns their actual results. For the April 14, 2025 Nano snapshot, OpenAI specifically recommends disabling parallel tool calls because the model can request the same tool several times in one turn. This is a snapshot-specific execution consideration, not a reason to interpret several proposed calls as completed actions.

Image input adds visual evidence to a text task, such as reading a visible label or describing the state shown in a screenshot. Use supported PNG, JPEG, WebP or non-animated GIF images, keep text legible and explain which observation matters. The vision guide notes difficulties with small or rotated text, precise spatial relationships, object counts and some non-Latin writing. Give surrounding context where a crop alone would make a label or diagram ambiguous. Nano’s own response is text; an image-generation tool returns its own separate artifact. 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.

In the producer’s file workflow, a vision-capable model receives both extracted text and page images from PDFs. Other document formats contribute extracted text, and spreadsheet augmentation considers up to the first 1,000 rows per sheet. A file must be under 50 MB, with 50 MB combined across the request. Chat Completions supports PDF file parts, while the documented Responses workflow accepts additional file types. Embedded charts in non-PDF documents are not passed as image content, so a PDF representation is useful when the visual evidence matters.

OpenAI documents prompt caching for Nano and Predicted Outputs for the GPT-4.1 Nano series. Prompt caching reuses an eligible input prefix in the producer service. Predicted Outputs instead supplies largely known output text for an edit, such as revising a document or code file while preserving its unchanged portions. The prediction path is text-only and excludes function calling, logprobs and max_completion_tokens, among other controls. Choose the path for the actual operation and keep its result associated with the original source.

The producer also documents supervised fine-tuning and direct preference optimisation for Nano’s named GPT-4.1 snapshot. Training examples and preferred-versus-rejected responses adapt a model to a recurring task; they are distinct from the context supplied to an ordinary request. For a practical extraction or classification stage, start with clear categories, examples and an explicit final format, then inspect whether the returned fields preserve the facts that the next stage will use.