OpenAI: GPT-4o-mini Search Preview

Executes real-time web searches within chat completion requests by routing queries through a search integration layer that retrieves current web results and injects them into the model's context window before generation. The model is fine-tuned to understand search intent signals in user prompts and automatically determine when web search is necessary versus when cached knowledge suffices, reducing unnecessary API calls while maintaining factual accuracy on time-sensitive queries.

The model internally classifies incoming queries to determine whether web search is required or if existing knowledge is sufficient, using learned patterns from training data to identify temporal signals (dates, 'latest', 'current'), factual domains (news, prices, events), and explicit search indicators. This routing decision happens before search execution, allowing the model to skip unnecessary searches and preserve context window tokens for queries answerable from training data.

Integrates web search results into the model's context window by formatting retrieved pages, snippets, and metadata into structured chunks that fit within token limits while preserving relevance ranking. The injection mechanism prioritizes high-relevance results and compresses verbose content to maximize space for user history and multi-turn conversation context, using a learned compression strategy to balance result fidelity with context availability.

Grounds model responses in real-time web data by retrieving current facts and enabling the model to cite sources directly from search results, reducing hallucinations on time-sensitive queries. The model is trained to recognize when citations are appropriate and to reference specific URLs, publication dates, or snippet text from search results, providing transparency about information provenance and allowing users to verify claims.

Maintains conversation history across multiple turns while selectively augmenting individual user messages with web search results, allowing the model to reference earlier context and build on previous responses while incorporating real-time data. The model tracks conversation state and determines which turns require search augmentation, avoiding redundant searches for follow-up questions that can be answered from earlier search results or training knowledge.

Integrates with OpenAI's Chat Completions API using standard request/response formats, supporting all Chat Completions parameters (temperature, max_tokens, top_p, etc.) while transparently handling search augmentation in the backend. The model accepts standard chat message arrays and returns responses in the same format as other GPT models, with optional metadata indicating search was performed, enabling drop-in replacement for existing Chat Completions workflows.