Meta has launched Muse Ember, an AI model with a 1M-token context window aimed at repeated coding, document processing and tool tasks. Available today through Muse Code and Meta Model API, it costs $0.10 per million input tokens and $0.50 per million output tokens. Meta describes it as the smallest model in its lineup and says it is designed for the many individual steps that agents perform during larger jobs.
Cached input costs $0.004 per million tokens, and the model can produce up to 512K output tokens. Developers can access it as muse-ember-1.0, with support for the Responses API. Meta says the context capacity lets developers submit an entire codebase or a collection of documents in one request.
Muse Ember accepts video, audio, images and text, which Meta says lets agents handle material such as meeting recordings, scanned pages and screenshots without switching to another model. The company also says it can control a computer using screenshots and identify small visual details. For business workflows, Meta describes agents retrieving records, making successive tool calls and changing information across applications.
Meta positions Muse Ember below Muse Wonder in its division of work: Muse Wonder can plan difficult jobs while Muse Ember executes the repeated steps. Coding uses include examining files, executing tests and applying focused changes. Muse Ember is also the first model trained using Muse Wonder Full, according to Meta.
Meta says Muse Ember matches the strongest Claude Haiku 5.5 score at roughly one-third of the cost and gives more intelligence for the price than GPT-6 Luna at every reasoning setting in its comparison. Those comparisons use Artificial Analysis intelligence scores and task costs, with Muse Ember costs estimated using proposed pricing. The company puts Muse Ember's cost at about $0.14 for each Artificial Analysis task, a separate measure from its token rates.
The launch follows Meta's move to bring Active Memory and Active Focus to specialized models serving Muse and Meta AI. For that earlier update, Meta reported a 67% gain in Muse's completion of everyday tasks involving unexpected obstacles, without specifying a rollout schedule.
Meta says its pre-release assessments placed Muse Ember below the high-risk thresholds in its Meta Superintelligence Scaling Framework for the categories tested. It also tested whether the model follows user intent, protect sensitive information, resist hidden instructions in external material and request human input before consequential actions.
