If you run an agent that makes hundreds of small calls an hour, the model behind those calls decides your bill more than any flagship does. That is the gap Meta is aiming at with Muse Ember, which it describes as its smallest and most efficient model, built for fast agentic execution at scale.
What Muse Ember is built for
Meta positions Ember for coding tasks, tool calls and document analysis: the repetitive steps inside an agent workflow rather than the hard planning. In Meta’s own example setup, the larger Muse Wonder acts as the planner while Ember carries out the individual steps. An agentic workflow, in this sense, is a chain where a model plans, calls tools, checks results and moves on without a person approving each step.
The model accepts text, image, audio and video input and offers a context window of 1 million tokens. A token is a small chunk of text the model reads; 1 million tokens is enough for very long documents or large code bases in one request.
What it costs and where you can use it
Up to 200,000 input tokens, Meta charges $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004. Above 200,000 input tokens the rates triple to $0.30, $1.50 and $0.012. The model is available through the Meta Model API under the ID muse-ember-1.0, works with OpenAI-compatible clients by changing the base URL, and is included in all Muse Code subscription plans.
How it compares with Claude Haiku 5.5 and GPT-6 Luna
The independent benchmarking firm Artificial Analysis measured a score of 36 on its Intelligence Index at maximum reasoning effort, the same as Claude Haiku 5.5 (max). The difference is cost: about $0.14 per index task for Ember against $0.41 for Haiku, partly because Ember used roughly a third as many output tokens. Meta lists GPT-6 Luna (max) at 28 on the same index.
The picture is not uniform. Ember leads Haiku 5.5 on Terminal-Bench 4.0 (40.9 versus 32.8 percent) and reaches 68.1 percent on AutomationBench-AA, but it trails on FrontierCode 2.0 (27.2 versus 34.9 percent) and on AA-Briefcase. Inside Muse Code it scored 46 on the Coding Agent Index, ahead of Haiku 5.5 in Claude Code (37) and GPT-6 Luna in Codex (31).
The catches worth knowing
Knowledge is the weak spot. On AA-Omniscience, Ember answered only 21.1 percent of questions correctly, the lowest among similarly capable models, compared with 34.7 percent for Haiku 5.5. It compensates by declining 64 percent of questions, which keeps its hallucination rate at 17.4 percent, about half of Haiku’s. The second catch is the price step above 200,000 input tokens: long-context jobs cost three times as much per token.
Where Ember fits
Start with the job you need done. If your pipeline routes, extracts or runs tool calls at high volume, these numbers make Ember a sensible default worker, and pairing it with a stronger planner matches how Meta itself frames it. If you need factual recall or the hardest coding work, the same measurements point the other way. My reading is that the refusal behaviour is a feature for agents, since a model that says it does not know is easier to route around than one that guesses, but it also means you should keep a retrieval step or a larger model for knowledge-heavy questions. Most of these figures come from launch-day testing, so independent results over the coming weeks will matter.
