Meta’s Muse Ember targets the repeated work behind agent…

The lightweight model offers cheaper execution for repetitive tasks, but benchmark savings are not evidence of better customer-support or sales outcomes.

Meta released Muse Ember on November 9, describing it as its smallest and most efficient model, designed for repeated agent tasks including tool calls, routing and data extraction.

Through the Meta Model API, the company lists prices of $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. Longer prompts cost $0.30 per million input tokens and $1.50 per million output tokens. The supported context window is roughly one million tokens.

Meta’s Artificial Analysis Intelligence Index comparison puts Ember at about $0.14 per task, against $2.39 for Muse Spark 1.3, with scores of 36 and 37 respectively. Those task costs use list pricing. They aren’t measurements of what it costs to resolve a customer inquiry.

Meta recommends pairing Ember with Muse Wonder for harder problems. TechCrunch illustrates that division with a hypothetical customer-support agent: a heavier model establishes handling rules, while a lighter model applies them across many cases. That is an architectural example, not a documented support deployment.

For an owner or founder running a small B2B team, the potential cost implication is narrower than a claim of better service. Separating difficult decisions from repeatable execution could make inquiry handling and follow-ups less expensive. That inference needs testing against the team’s actual workload; the release doesn’t establish improved lead generation, sales or customer-support outcomes.

Access also comes with provider-specific limits. Vercel says Ember is available through AI Gateway with adjustable reasoning effort. Google Cloud lists Muse Ember 1.0 on Gemini Enterprise Agent Platform as a preview supporting text and image inputs, function calling and structured output, but not batch predictions. Meta describes broader native media capabilities. Availability through one provider shouldn’t be treated as identical to another.

Nor does Meta’s benchmark evidence show Ember ahead on every task. Its published FrontierCode 2.0-AA results put Claude Haiku 5.5 at 34.9%, against Ember’s 27.2%. The Decoder also highlighted that gap.

For businesses deploying agents to serve customers, the buying question is whether cheaper repeated steps can complete their customer-facing workflow reliably enough to justify deployment. The benchmark comparison supports examining execution costs, not assuming that lower token bills will translate into better business results.

Meta’s Muse Ember targets the repeated work behind agent costs

Sources