o4-mini-deep-research

What is o4-mini-deep-research?

o4-mini-deep-research is OpenAI’s specialized reasoning model for research tasks that require several stages of work rather than a short, direct answer. A typical workflow might involve searching for relevant sources, extracting information, comparing conflicting claims, analyzing attached documents, and writing a report that connects the evidence.

The model is designed for this type of investigation rather than for every possible AI workload. OpenAI describes it as a faster, more affordable deep research model. In practical terms, it occupies a middle position: it is more focused on sustained research and source synthesis than a general conversational model, while aiming to cost less and complete tasks faster than OpenAI’s larger o3-deep-research option.

The current canonical model alias is o4-mini-deep-research. OpenAI also lists o4-mini-deep-research-2025-06-26 as a dated snapshot, but that snapshot is marked deprecated. Applications should therefore use the canonical alias when they need the currently documented model identity.

How its research workflow works

o4-mini-deep-research is intended for investigations where the answer cannot be produced reliably from a single prompt and the model’s existing knowledge. Through the Responses API, it can work with several external tools:

• Web search: retrieves current information from online sources during a research run.
• Remote MCP servers: connects to compatible external search and fetch services. Deep research MCP integrations need tools that expose the required search and retrieval operations.
• File search: searches documents stored in vector stores, allowing a workflow to use an organization’s internal material.
• Code interpreter: supports data analysis and computation as part of a research process.

These tools do not change the model’s underlying training cutoff. OpenAI lists the knowledge cutoff as June 1, 2024, while web search and connected sources can provide newer information during an individual run. That distinction matters: the model can research events after its cutoff when the appropriate tools are enabled, but an offline prompt should not be assumed to contain those later facts.

Inputs, modalities, and output

The verified modality profile is relatively focused. The model accepts text and image inputs and produces text output. Images can therefore be included as evidence or context for an investigation, but the model is not an image-generation system.

• Text input: supported.
• Image input: supported.
• Audio input: not supported.
• Video input: not supported.
• Text output: supported.
• Image, audio, or video output: not supported.

This makes o4-mini-deep-research appropriate for written research deliverables such as reports, evidence summaries, market analyses, and document reviews. It is not the right choice for voice conversations, media generation, or applications that require the model itself to return images, audio, or video.

Context window and technical limits

o4-mini-deep-research has a 200,000-token context window and supports a maximum output of 100,000 tokens. A token is a unit of text processed by the model; the exact number of tokens in a document depends on the language and formatting. These limits are large enough for workflows involving substantial source material, long files, and detailed final reports, although the available space must still be shared between instructions, retrieved content, conversation history, and generated output.

OpenAI’s model documentation lists function calling and structured outputs as unsupported. This is an important practical limitation. Although the model can use research tools through the documented deep research workflow, that should not be confused with support for a conventional function-call interface or schema-constrained JSON responses. Applications that require those features may need another model or an orchestration layer outside the model.

Pricing and cost trade-offs

The documented standard pricing is $2.00 per 1 million input tokens, with cached input priced at $0.50 per 1 million tokens. Output costs $8.00 per 1 million tokens. The input and output rates apply to different parts of a request, so a long generated report can cost more than a short answer even when the source prompt is unchanged.

OpenAI supports batch processing for this model, but the supplied model documentation does not verify an exact batch price. The safest interpretation is to use the standard rates above for ordinary requests and treat batch availability as a separate operational feature rather than assuming a particular discount.

The main value proposition is not simply a low per-token price. Deep research tasks can involve long prompts, retrieved sources, tool activity, and lengthy reports, so total cost depends on the complete workflow. o4-mini-deep-research is attractive when the task needs substantial reasoning and source synthesis but does not justify the cost or slower execution associated with a larger deep research model.

Reasoning and coding profile

Its reasoning capability is central to the model’s purpose. The model is intended to break down complex questions, work across multiple sources, compare evidence, and synthesize a coherent answer. This is different from simply retrieving a passage or completing a short prompt: the useful result is the chain of investigation that leads to a written conclusion.

Code interpreter support also makes the model useful for research involving structured data, calculations, or transformations that are difficult to perform reliably by hand. For example, a workflow could combine retrieved market information with calculations or analyze a supplied dataset before presenting the findings. The supplied research does not establish that o4-mini-deep-research is a dedicated software-engineering model, however. Its coding support is best understood as a research and data-analysis tool rather than a guarantee of specialized programming performance.

The editorial research profile rates its reasoning at 9 out of 10, coding at 7 out of 10, speed at 8 out of 10, and cost at 8 out of 10. These are editorial evaluations, not OpenAI-published benchmark scores or official performance guarantees. They summarize the model’s intended balance and should not be treated as standardized test results.

Best use cases

o4-mini-deep-research is a strong fit when the task requires both investigation and explanation. Suitable examples include:

• Preparing legal or scientific research summaries from multiple sources.
• Comparing products, companies, markets, or competing claims.
• Analyzing large collections of internal documents through file search.
• Building evidence-based reports that require current web information.
• Investigating a question that involves several related subquestions.
• Combining retrieved information with calculations or data analysis.
• Reviewing images alongside written documents as part of a broader research task.

The model is particularly useful when the final deliverable is a detailed written explanation rather than a single fact. Its long context and large output ceiling also help with source-heavy assignments, provided the application controls the amount and quality of retrieved material.

When should you choose o4-mini-deep-research?

Choose o4-mini-deep-research when you need a research-oriented reasoning model with external information access, but want a faster and less expensive option than OpenAI’s larger deep research model. It is a practical choice for recurring investigations, internal knowledge review, market research, and reports where the quality of source synthesis matters more than generating non-text media.

A conventional general-purpose model may be more suitable for short questions, routine drafting, or workflows that do not need web research and multi-step analysis. A model with ordinary function calling or structured-output support may be preferable when the application must return data that conforms to a strict schema. A voice or media model is a better fit for real-time audio, image generation, video, or other non-text output requirements.

Within research workloads, the choice is mainly a capability-versus-cost decision. o4-mini-deep-research is intended to preserve the core deep research workflow while reducing cost and improving speed relative to the larger o3-deep-research model. The trade-off is that it remains specialized: it should not be selected merely because it has a large context window if the application’s primary need is structured automation, fine-tuning, media generation, or highly specialized software engineering.

Limitations to plan for

Several constraints should be considered before deployment:

• The model produces text only, so downstream media generation requires another system.
• Audio and video inputs are not supported.
• Function calling and structured outputs are listed as unsupported, limiting direct integration with strict tool schemas and machine-readable response contracts.
• Fine-tuning is not supported.
• The June 1, 2024 knowledge cutoff means current information requires web search or another connected source.
• Tool-enabled research can introduce source-quality, retrieval, and orchestration considerations that are separate from the model’s reasoning ability.

Overall, o4-mini-deep-research is best understood as a focused research engine: it is built to investigate, analyze, and write, with external tools extending what it can access. Its appeal comes from the combination of long context, substantial output capacity, tool-assisted research, and a lower-cost positioning within OpenAI’s deep research lineup—not from broad support for every type of AI application.