Models

Scout and Maverick ship amid benchmark controversy and a missing flagship.

Reasoning models gain the ability to use tools inside their chain of thought.

The model spends extra computation thinking before it answers.

Meta releases a downloadable model at roughly GPT-4 class.

Voice and vision at frontier quality open to all users at no cost.

Opus posts the first credible benchmark wins over GPT-4.

Text prompts generate near-photorealistic minute-long video.

Whole codebases and books fit in a single prompt.

A natively multimodal model family is unveiled across three sizes.