Meta Muse Ember: a ten-cent model for the grunt work agents do all day
Ten cents per million input tokens, fifty cents per million output. That is the price of Muse Ember, which Meta released Monday as its smallest and cheapest Muse model, built for what Meta calls "the work agents repeat all day": tool calls, coding subtasks, routing, extraction, and long document parsing.
On Meta's own table, against Claude Haiku 5.5 and GPT-6 Luna at max reasoning, the agent scores are the eye-catcher. AutomationBench-AA, end-to-end business workflows on strict pass@1, is 37.8% for Ember against 13.9% for Haiku and 20.2% for Luna. Terminal-Bench 4.0 is 40.9 vs 39.2 vs 16.4. Coding Agent Index-AA is 46 vs 37 vs 31. Then the loss that Meta also prints: FrontierCode 2.0-AA, real engineering tasks judged on mergeability, has Haiku at 34.9 and Ember at 27.2. On the overall Artificial Analysis Index, Ember and Haiku tie at 36.
The cost claim is the real pitch. Meta says Ember matches Haiku's best score at about a third of the cost per task, $0.136 against $0.411, and sits on the efficiency frontier of the whole Artificial Analysis board. Caveat in Meta's own fine print: Ember's costs are "estimates at proposed pricing". Two of the benchmarks, Agents' Last Exam and the Personal Superintelligence Index, are Meta's internal ones. It has a 1M-token context and up to 512K output tokens, and is live on the Meta Model API.
The timing makes the story. On the same day Anthropic pitched Fable 5.2 as an Orchestrator that hands simple tasks to smaller models, Meta shipped a model whose whole job is to be the smaller model. Ember is meant to sit under Muse Wonder, which does the planning. Agent harnesses are turning into org charts: an expensive planner up top, cheap workers doing the volume. Whoever owns the cheap-worker tier gets most of the tokens, and Meta just priced in hard. The FrontierCode gap says to keep the merge-quality work on a bigger model for now.
Source: https://research.meta.ai/blog/muse-ember and the evaluation report https://research.meta.ai/static/muse-ember-evaluation-report
← Back to all articles
On Meta's own table, against Claude Haiku 5.5 and GPT-6 Luna at max reasoning, the agent scores are the eye-catcher. AutomationBench-AA, end-to-end business workflows on strict pass@1, is 37.8% for Ember against 13.9% for Haiku and 20.2% for Luna. Terminal-Bench 4.0 is 40.9 vs 39.2 vs 16.4. Coding Agent Index-AA is 46 vs 37 vs 31. Then the loss that Meta also prints: FrontierCode 2.0-AA, real engineering tasks judged on mergeability, has Haiku at 34.9 and Ember at 27.2. On the overall Artificial Analysis Index, Ember and Haiku tie at 36.
The cost claim is the real pitch. Meta says Ember matches Haiku's best score at about a third of the cost per task, $0.136 against $0.411, and sits on the efficiency frontier of the whole Artificial Analysis board. Caveat in Meta's own fine print: Ember's costs are "estimates at proposed pricing". Two of the benchmarks, Agents' Last Exam and the Personal Superintelligence Index, are Meta's internal ones. It has a 1M-token context and up to 512K output tokens, and is live on the Meta Model API.
The timing makes the story. On the same day Anthropic pitched Fable 5.2 as an Orchestrator that hands simple tasks to smaller models, Meta shipped a model whose whole job is to be the smaller model. Ember is meant to sit under Muse Wonder, which does the planning. Agent harnesses are turning into org charts: an expensive planner up top, cheap workers doing the volume. Whoever owns the cheap-worker tier gets most of the tokens, and Meta just priced in hard. The FrontierCode gap says to keep the merge-quality work on a bigger model for now.
Source: https://research.meta.ai/blog/muse-ember and the evaluation report https://research.meta.ai/static/muse-ember-evaluation-report
Comments