Meta announced last Monday (Nov 9) Muse Ember, the lightest and most cost-efficient AI model in its Muse lineup. Positioned as the high-volume execution layer within a multi-model system — paired with the larger Muse Wonder for planning — Ember arrives at $0.10 per million input tokens, with an estimated $0.14 per task on the composite metric from Artificial Analysis. Numbers the company says place Ember at the general leaderboard efficiency frontier, below both Claude Haiku 5.5 and GPT-6 Luna in cost per task, even with comparable scores.
What makes Ember different is not raw capability, but architectural positioning. Meta describes it as built specifically for complex reasoning execution and the work agents repeat all day: sequential tool calls, coding subtasks, routing, information extraction, and long document analysis. It's the same principle that motivated sub-agent frameworks from companies like Anthropic and OpenAI, but with one crucial difference: Ember is not a generic low-capability model, it's an execution specialist.
The math of cost per task
The core of the Ember launch is the cost-per-task metric, not just price per token. Language models are billed per input and output token, but the true cost of an autonomous agent lies in how many tokens it spends to solve a complete problem. Artificial Analysis measures this with the Intelligence Index, which evaluates the model's composite performance across a battery of tests, then divides the total cost per task.
By that metric, Ember scores 36 on the index at $0.14 per task. Claude Haiku 5.5 also scores 36 at $0.41 — almost three times more expensive. GPT-6 Luna, meanwhile, costs $0.15 per task but scores only 28 on the same index. The crucial difference between the three models that cost the same per token — $0.10 input and $0.50 output per million tokens — lies in cache tokens: Ember charges $0.004 per million, while Haiku and Luna charge about 60% more. Since cache represents the bulk of cost in agentic evaluations, that difference translates into substantial savings.
In practice, this means an agent configured to process thousands of support calls per day can use Ember to classify, extract data, and route requests for pennies, while Wonder — or another frontier model — handles only the complex decisions that genuinely require deep planning. Meta illustrated this with a customer support example: Wonder defines the rules for handling each type of issue, and Ember applies those rules across thousands of cases.
Benchmarking: where Ember shines (and where it doesn't)
In Meta's published tests, Ember dominated agent-oriented tasks. On AutomationBench-AA, which measures end-to-end enterprise workflows, Ember scored 37.8%, compared to Haiku's 13.9% and Luna's 20.2%. On Agents' Last Exam — which evaluates the model's ability to execute real agentic tasks in a functional harness — Ember scored 61.6%, ahead of Haiku's 57.0% and Luna's 57.1%. On Terminal-Bench 4.0, which evaluates terminal coding in an agentic harness, Ember reached 40.9%, ahead of Haiku (39.2%) and significantly ahead of Luna (16.4%).
But Ember doesn't win everything. On FrontierCode 2.0-AA — which judges mergeability of code on real engineering tasks — Haiku scored 34.9% against Ember's 27.2%. On Artificial Analysis's focus on hallucination rate, Ember answers 64% of questions with a 17.4% hallucination rate, half of Haiku's rate, suggesting that Ember tends to be more conservative in answer generation.
This mix of strengths and weaknesses is typical of execution models: they shine when work is structured and repetitive, but lose to larger models on tasks requiring creativity or complex code sophistication. Meta knows this, which is exactly why Ember was not designed to operate alone.
The Muse ecosystem as market strategy
Ember is the fourth model in Meta's Muse line, following Spark (launched in April, mid-weight model for agents and coding), Glimmer (open-source twin of Spark, in August), and Wonder (frontier model, in October). This progression of capability — from consumer (Muse, launched in September with over 10 million downloads in the US within the first 40 days) to frontier (Wonder), middle-tier (Spark), to execution (Ember) — reflects a deliberate strategy to create a complete agentic reasoning pipeline.
Meta trained Ember using distillation from Muse Wonder Full, a process in which a larger model generates or evaluates training signals for the smaller model to specialize. This places Ember in the teacher-student model category — meaning its performance is not limited by weight size, but by the quality of the teacher. If Wonder Full can generate high-quality signals, Ember can achieve performance far above what a model trained from scratch would reach at the same size.
Ember's availability comes through the Meta Model API and Muse Code, with Responses API support. It is also available via OpenRouter as Crouton Alpha (formerly a free beta version of Ember). Meta has not published Ember's weights — as it does with Spark and Wonder — keeping it closed, which limits self-hosted infrastructure adoption, but maintains control over model quality and security.
What this means for the agent market
The Ember launch signals a phase shift in the AI market: the race is no longer about ever-larger models, but about models that are increasingly efficient at executing repetitive tasks. The multi-model architecture — where one large model plans and one small model executes — is already well known in the autonomous agent community, but Ember takes this idea to the commercial level with a price that makes scale execution economically viable for companies of all sizes.
If cost per task drops to cents, AI agents stop being lab prototypes and become product components. The question is no longer how many tasks can an agent perform, but how many tasks can it perform without breaking the budget. And in that sense, Ember is telling the market that the answer is a lot.
Sources: TechCrunch, Meta Research, Artificial Analysis
✓ Independent sources cross-checked and verified before publishing
