Can Chat Completions search models answer without searching?

The Citead team, updated

In short. Chat Completions search models always retrieve web information before responding, while Responses API search can be optional.

Definition: No. OpenAI documents that the specialized search models used with Chat Completions retrieve information from the web before responding. Optional web search is documented for the Responses API, where the model can choose whether to use the web_search tool.

Key facts

  • OpenAI's web search documentation says Chat Completions supports specialized search models for web search.
  • The same guide says a Chat Completions search model always retrieves information from the web before responding.
  • The guide lists gpt-5-search-api for the current Chat Completions search model path.
  • The legacy gpt-4o-search-preview and gpt-4o-mini-search-preview models are marked deprecated and scheduled to shut down on 2026-07-23.
  • Chat Completions responses include inline citations for URLs found in web search results, with URL and title data in a url_citation annotation.

What does the specialized completion endpoint do?

Chat Completions accepts a list of messages and returns a model response. For web search, OpenAI documents a specialized model path rather than a general model with an optional hosted search tool. The current guide lists gpt-5-search-api; it also lists the two preview search models as deprecated and gives 2026-07-23 as their shutdown date.

The request enables the search path with web_search_options. The documented example leaves that object empty, so the request does not contain a switch that turns the search step off:

{
  "model": "gpt-5-search-api",
  "web_search_options": {},
  "messages": [
    {"role": "user", "content": "What was a positive news story from today?"}
  ]
}

Can the model choose to skip the web?

The answer depends on the API path. OpenAI states that, with Chat Completions, the model always retrieves information from the web before responding to the query. The same guide describes Responses API web search as a tool that the model can choose to use based on the input prompt. With tool_choice: "auto", search is optional in that path; tool_choice: "required" or a specific web search choice makes search run.

This difference is specific to the documented interfaces. The Chat Completions search models do not support several Responses API search controls, including domain filters, complete source lists, live-access control and returned-token budget control.

What does the response contain?

OpenAI says the Chat Completions response contains text in message.content, including inline citations. The web search guide says the url_citation annotation includes the cited URL, its title and the start and end character indexes where the source is used. The Chat Completions reference describes a URL citation as an annotation used with web search.

{
  "type": "url_citation",
  "url_citation": {
    "end_index": 985,
    "start_index": 764,
    "title": "Page title...",
    "url": "https://..."
  }
}

OpenAI also says that inline citations must be clearly visible and clickable when an application displays web results or information from them to end users.

FAQ

Does an empty web_search_options object make search optional?

No. The documented Chat Completions search request uses an empty object, while OpenAI states that Chat Completions search models retrieve web information before responding.

The Responses API supports the hosted web_search tool. With tool_choice: "auto", OpenAI documents that the model can choose whether to search.

Do the preview Chat Completions search models remain available?

OpenAI marks gpt-4o-search-preview and gpt-4o-mini-search-preview as deprecated and gives 2026-07-23 as their shutdown date. The guide lists gpt-5-search-api for the Chat Completions search model path.

Sources

  1. OpenAI Developers, "Web search", undated
  2. OpenAI Developers, "Chat | OpenAI API Reference", undated

Check how a page reads to answer engines. Open the GEO score analyzer

Test a page and see what to fix first

The analysis usually takes less than 10 seconds and needs no account.

400,000+ pages scored in the public benchmark