Meta Muse Ember Priced at 10 Cents per Million Input Tokens

By Rebecca Stone

Meta Superintelligence Lab released Muse Ember on Monday at 10 cents per million input tokens, a rate Meta says makes the model its most cost-efficient AI system to date.

Ember is also the lab’s lightest model. Meta described it as “designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.”

Volume changes the economics.

What does 10 cents per million tokens change?

Meta’s 10-cent input-token rate places Ember significantly below Anthropic’s Haiku and OpenAI’s Luna in the lab’s comparison. The company cited Artificial Analysis’s composite benchmark score to argue that Ember is both cheaper and smarter than the two competitors.

Meta also called Ember cost-competitive with open-source DeepSeek V4.1 Flash and GLM 5.3 Flash. Thus, the new model targets a different price band from frontier systems without surrendering its claim against lower-cost open models.

The 10-cent figure is a disclosed input-token rate, not an all-in cost per completed task. The release gives no output-token rate, and its cost-per-task claim rests on Meta’s comparison rather than an independently reported customer bill.

How does Ember work with Wonder?

Meta designed Ember to sit beside the larger Muse Wonder in an agentic harness. In the customer-support example, Wonder sets the rules for a particular issue, while Ember applies those rules across tens of thousands of cases.

“Muse Wonder plans, Muse Ember handles the volume,” Meta wrote. The architecture separates policy and planning from repetitive execution, potentially allowing a company to reserve its most expensive model calls for exceptions.

That pairing remains a proposed deployment pattern in the announcement, not a disclosed rollout. Meta named neither customers nor a deployment count, and it published no benchmark showing how much Wonder-led routing lowers total system cost.

Can one model handle mixed media?

Meta trained Ember across text, images and audio. The company says agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video without a separate model.

That claim broadens Ember’s addressable work beyond text-only coding or routing. It also positions the 10-cent model as a common execution layer for support, document processing and meeting-related tasks, although Meta disclosed no modality-specific accuracy, latency or pricing figures.

Where does Ember fit in the Muse lineup?

Ember is the fourth model under Meta’s Muse brand. The release sequence and positions are:

• Muse Spark, a mid-weight model, released in April.
• Glimmer, Spark’s open-source twin, released in August.
• Muse Wonder, the frontier model, released in October.
• Muse Ember, the lightweight, cost-focused model, released Monday.

Separately, Meta released the flagship consumer product, Muse Agent, in September, which the company said had achieved unexpected success. Its Spaces and video-chat features began rolling out earlier this month, but Meta supplied no user, revenue or engagement figures for either release.

Ember therefore extends a lineup already spanning consumer software, open-source and frontier models. It also gives Meta a lower-cost execution option for repeated agent work.

The commercial test is whether that execution option can scale without eroding accuracy: a 10-cent input rate can undercut Haiku and Luna on high-volume jobs, while cost-competitive open models such as DeepSeek V4.1 Flash and GLM 5.3 Flash leave Meta little room to move lower. Meta must now show whether real agent traffic can validate the price advantage while keeping quality consistent across Wonder’s planning and Ember’s execution.