OpenAI: o3 (batch)

o3 is OpenAI’s general-purpose o-series reasoning model for complex tasks. The manufacturer highlights mathematics, science, coding, visual reasoning, technical writing and instruction following. It accepts text and images and produces text, so a task can combine a written problem, code and a diagram rather than treating each source as an isolated question. The intended result is a considered answer to a multi-step problem, with the information required to assess that answer.

For a mathematical or scientific question, provide the complete problem, definitions, known quantities, units and constraints. State whether you need a derivation, a numerical result, a comparison of hypotheses or an explanation for a particular audience. An illustrative output brief is: identify the relevant assumptions, present the solution with its units, and explain which supplied facts support it. Ask for a useful explanation in the final answer; OpenAI’s internal reasoning tokens are not exposed as the model’s raw working trace.

For software work, supply the relevant code, the expected behavior and the observed failure, including the exact error or input that reveals it. Ask o3 to connect the symptoms to a plausible cause and propose a focused change with an explanation. A visual task can include a diagram, screenshot or plotted result: identify the question and the labels or region that matter, retain a legible scale, and request a conclusion tied to what is actually visible. These preparation steps give the model a defined problem and a concrete output instead of an unrestricted request to inspect everything.

OpenAI lists a 200,000-token context window, a maximum output of 100,000 tokens and a knowledge cutoff of June 1, 2024. Internal reasoning uses context and output capacity alongside the visible answer. The maximum output is therefore not a promise of 100,000 visible answer tokens, and the context is not an additional independent output allowance. Leave room after the input material for analysis and the requested result. The producer’s reasoning guide describes incomplete responses when a context or output limit is reached; this can happen before a visible answer appears.

The exact o3 specification documents Chat Completions, Responses and Batch interfaces, together with streaming, Structured Outputs, function calling, file uploads, image input, file search and prompt caching. Its documented tool set includes web and file search, code interpretation, image generation and MCP connections. A search result or calculation is additional evidence for the task, while a function call requests an operation from the surrounding application. Keep the returned evidence associated with the question it answers, and distinguish a proposed action from a confirmed result. The model’s own output remains text even when an image-generation tool creates a separate artifact.

OpenAI’s reasoning guidance favors simple, direct instructions with clear boundaries between instructions and source material. State the goal, constraints and required answer format instead of prescribing the model’s private thought process. Use examples when they clarify a difficult format or decision rule, and keep those examples consistent with the actual requirements. For a long technical document, label sections and ask for the passages or relationships relevant to a specific question; for code review, specify the behavior and risk you want investigated.

A useful o3 deliverable can be a worked technical explanation, an analysis of a diagram, a documented code correction or a comparison grounded in the supplied material. Make unsupported assumptions and missing information explicit, and check decisive calculations or code behavior against the original problem. Its knowledge cutoff means that recent facts need supplied material or an appropriate information tool. Clear evidence, an explicit completion criterion and sufficient output space help turn a broad reasoning task into a result that another person can inspect and use.