Meta introduced Muse Ember on Monday, Nov. 9, for repeated AI-agent tasks, with lower API rates on requests up to 200,000 input tokens and uses including tool calls and coding subtasks.
Meta's Model API charges $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens for requests at or below that threshold. Above it, rates rise to $0.30, $0.012 and $1.50, respectively, with the higher prices applying to the entire request. The input and output prices match OpenAI's listed rates for GPT-6 Luna and Anthropic's rates for Claude Haiku 5.5 prompts up to 100,000 tokens. Meta also offers Ember through Muse Code and says it handles routing, extraction and long-document parsing.
Meta's API listing identifies the model as muse-ember-1.0, with a context window of 1,048,576 tokens and a maximum output of 524,288 tokens. Meta says agent workflows involve many repeated small steps, a rationale for the model's focus on frequent calls. Developers using requests above the input threshold face higher prices across the full request.
Meta's launch chart, drawing on Artificial Analysis Intelligence Index v5 results dated Nov. 9, lists Ember at 35.8 in maximum-reasoning mode, with an estimated cost of $0.136 per task. Claude Haiku 5.5 scored 36.2 at $0.411 per task, while GPT-6 Luna scored 27.8 at $0.152.
Results varied across other tests. Meta reported Ember at 37.8% on AutomationBench-AA strict pass@1, compared with 20.2% for Luna and 13.9% for Haiku. On FrontierCode 2.0-AA, Haiku scored 34.9%, ahead of Ember at 27.2% and Luna at 25.2%.
Meta says Ember is the first model trained using Muse Wonder Full. The company describes Wonder as suited to harder problems and Ember as a model for repeatable coding and agent tasks at scale.
Meta quoted Cloudflare Chief Technology Officer Dane Knecht as saying early tests found Ember was “up to 30% faster at the tail” than a comparable model Cloudflare tested. Meta's announcement provided no further testing details.
