Token-by-token responses streamed as they generate.

Call developer-defined functions from the model.

Constrain output to valid JSON.

Return responses matching a supplied schema.

Understand and reason over images.

Accept images as part of the prompt.

Accept and process audio input.

Generate images from text.

Understand and reason over video.

Extended step-by-step reasoning before answering.

Invoke tools and take agentic actions.

Native Model Context Protocol support.

Cache prompt prefixes to cut latency and cost.

Process large jobs asynchronously at lower cost.

Produce vector embeddings for search and RAG.