Meta has released Muse Ember for the repeated execution work inside AI agents, including tool calls, coding subtasks and document processing. The November 9 announcement positions it as a smaller model for workloads where many individual steps accumulate into substantial usage.
The model’s pricing page lists $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. Above that threshold, the listed prices rise to $0.30 and $1.50 respectively. Cached input costs $0.004 per million tokens in the lower tier and $0.012 in the higher one. The context window reaches one million tokens.
Meta says Ember is available through Muse Code and the Meta Model API under the identifier muse-ember-1.0, with Responses API support. It accepts text, image, audio and video inputs.
The intended deployment can involve multiple models. Meta’s product page describes pairing Ember’s repeated execution with larger Muse models handling planning and orchestration. That division gives developers another way to allocate work within a single agent workflow.
What the comparisons actually measure
Meta’s evaluation methodology report explains why the surrounding software matters. Its controlled internal comparisons use a common framework, but external coding results retain the agent or harness used by the reporting provider. The report also cautions that competitor prompts, tools and runtime settings may not be optimally tuned.
The published results do not show a uniform lead. Meta’s model page gives Ember 27.2% on FrontierCode 2.0-AA, behind Claude Haiku 5.5’s 34.9%. Other displayed tests favour Ember. These are benchmark-specific comparisons, rather than evidence that one model will perform best in every deployment.
For computer use, the methodology distinguishes partial task credit from complete success. OSWorld scores can therefore describe progress through a workflow without establishing that every required action finished correctly.
Meta says its safety assessment covered cybersecurity and loss-of-control risks, including resistance to instructions embedded in external content. Those are the company’s assessment claims. A lower token price and promising test results still leave developers with the practical task of checking completion quality, retries and operating cost in their own workflows.
