GPT-4o Mini Search Preview

A Search Model, Not Just GPT-4o Mini With Another Name

The important distinction was behavioral. Standard GPT-4o Mini is a general-purpose small multimodal model. GPT-4o Mini Search Preview was optimized around web search and current-information retrieval.

In the legacy Chat Completions search path, the model searched the web before responding. That made it useful for questions where the answer depended on information outside the model's October 2023 knowledge cutoff: recent news, current facts, changing product information, travel research, market context, and other time-sensitive topics.

OpenAI released the search preview models in March 2025 as direct Chat Completions access to the same family of web-search capabilities being introduced through the Responses API.

The model is now retired. OpenAI shut down the preview search models on July 23, 2026, so this page is most useful for understanding historical integrations and planning or documenting migrations.

• Specialized for web search through Chat Completions.
• Text-only input and text output.
• 128K context window and 16,384-token maximum output.
• Shut down on July 23, 2026.

Citations Were Part of the Product Contract

Search responses included inline citations and URL citation annotations. Those annotations exposed source metadata such as the cited URL and page title, allowing a client application to render citations alongside the generated answer.

This matters for product design. A search-backed answer should not look identical to an unsupported model response when sources are available. Users need a way to inspect where time-sensitive claims came from.

OpenAI's web search guidance explicitly requires citations to be clearly visible and clickable when web-derived information is shown to end users.

For a legacy migration, citation rendering should therefore be treated as a functional requirement rather than decorative UI. A replacement should preserve source visibility while taking advantage of current search controls.

Search Results Compete for Context Too

In a search application, context is not only the user's message and conversation history. Retrieved web material also needs to fit into the model's working context.

The legacy search interface exposed search_context_size, allowing developers to influence how much search-result context was provided to the model before answer generation.

A larger search context can expose more evidence, but it can also increase token usage and give the model more material to reconcile. A smaller search context can reduce cost and latency but may omit useful supporting information.

Current Responses API web search provides newer controls beyond the old specialized search-model path, including features such as filters, live-access control, and returned-token budgeting.

• Context window: 128,000 tokens.
• Maximum output: 16,384 tokens.
• Search-result material consumes part of the effective context.
• Search context size was an important quality, latency, and cost variable.

Prefer Responses API web_search for New Search Workflows

OpenAI's current migration guidance recommends the Responses API web_search tool for legacy GPT-4o search integrations.

This changes the architecture in a useful way. Instead of selecting a dedicated model that always searches, the application uses a current model with search as a hosted tool. The current search tool supports controls that the old Chat Completions preview models did not provide, including domain filters, external web-access control, and returned-token budgeting.

If an application must remain on Chat Completions, OpenAI points developers to gpt-5-search-api.

The migration should preserve more than answer text. Re-test citation rendering, current-information accuracy, search frequency, latency, query cost, and any structured-output contract the old integration relied on.