Back to news
launchMeta2026-11-10

Meta launches the lightweight Muse Ember model on November 10, matching Claude Haiku 5.5 on benchmarks at a third of the per-task cost

Muse Ember offers a one million token context window with up to 512 thousand output tokens, scores close to Claude Haiku 5.5 on Artificial Analysis, and is priced at 0.10 dollars per million input tokens and 0.50 dollars per million output tokens

On November 10, Meta released Muse Ember, the smallest and most efficient model the company has produced so far, designed for complex reasoning tasks and for the repetitive work agents handle day to day. Meta said the model matches Claude Haiku 5.5 on the Artificial Analysis leaderboard while costing roughly a third as much per task, and surpasses GPT-6 Luna on several tests.

Muse Ember has a one million token context window and can output up to 512 thousand tokens in a single call, so it can process a complete codebase or a full document set at once and run hundreds of programming subtasks within one session, including reading files, running tests and applying targeted fixes. The model natively handles text, images, audio and video, letting agents read screenshots, charts and scanned documents and process voice memos, meeting recordings and video without routing through another model.

On safety, the model was evaluated before release under Meta super intelligence expansion framework testing, with results showing a risk level below the high risk threshold set by that framework. Pricing is 0.10 dollars per million input tokens and 0.50 dollars per million output tokens, available from day one through the Meta model API platform, Google Cloud and Oracle Cloud.

Muse EmberMetaClaude Haiku 5.5轻量模型模型定价多模态