Meta Launches Muse Ember, Its Smallest AI Model, at $0.10 per Million Tokens

Built for agents

Muse Ember offers a context window of 1 million tokens and up to 512,000 output tokens, enough to load a whole codebase or document set in a single call. It also supports the Responses API. The model natively understands text, images, audio and video, so an agent can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video, without a separate model. It can also operate a computer by looking at screenshots.

Meta envisions it running workflows end to end: pulling records, calling tools in sequence and updating systems across several apps. In coding, it is suited to the hundreds of file reads, test runs and targeted fixes an agent makes in one job. For the hardest problems, Meta pictures pairing it with Muse Wonder: "Wonder plans, Ember does the volume." Ember is also the first model trained using Muse Wonder Full, the higher version of Wonder. Together with Muse Spark for coding and agent work, the lineup now has three tiers.

Cost claims and safety checks

Meta says that on Artificial Analysis's overall index versus cost per task, Ember delivers more performance per dollar than Claude Haiku 5.5 and GPT-6 Luna at every reasoning setting. It reportedly matches Haiku's top score at about one third of the cost and beats Luna in all settings. Cost per task is about $0.14, an estimate based on list prices.

Meta argues that delivering "personal superintelligence" to everyone requires frontier capability at a price that scales, and says even more efficient models are in training.

Before release, Meta assessed chemical and biological, cybersecurity and loss-of-control risks under its own framework, and all fell below the high-risk threshold. It also tested agent behavior, including whether the model follows instructions hidden in web pages or documents and whether it asks for human judgment before high-impact actions.