Meta's Muse Ember Undercuts Rivals at 10 Cents per Million Tokens

Meta's Superintelligence Lab put a hard number on its newest model this Monday: Muse Ember prices at 10 cents per million input tokens, making it the cheapest model the lab has shipped and, by Meta's own benchmark data, cheaper and smarter than Anthropic's Haiku and OpenAI's Luna on Artificial Analysis's composite score.

That price point matters because Ember isn't built to win poetry contests. It's built for the unglamorous work that AI agents repeat all day. Meta describes the model as "designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing."

For an enterprise buyer running thousands of automated workflows, the arithmetic is straightforward. A high-volume routing or extraction pipeline that chews through millions of tokens daily now has a frontier-lab option that competes not just with Anthropic and OpenAI's small models, but with open-source players too. Meta says Ember is cost-competitive with DeepSeek V4.1 Flash and GLM 5.3 Flash — pricing territory the closed-model vendors rarely enter.

What does Ember change for agent architectures?

The short answer: it formalizes the two-tier pattern most agentic systems already use informally.

Ember is designed to pair with a larger model, specifically Meta's Muse Wonder, inside an agentic harness. Picture a customer-support deployment. A heavier model establishes the rules for handling a particular issue, then hands the work to the lightweight model, which applies those rules across tens of thousands of cases. Meta's own framing in the announcement is blunt: "Muse Wonder plans, Muse Ember handles the volume."

That division of labor is the whole pitch. You pay frontier-model rates for the thinking and budget-model rates for the grinding — and the grinding is where token counts explode.

Multimodal by default

Ember also arrives trained across text, images and audio, which widens the range of inputs an agent can process without chaining in additional models. As Meta puts it in the announcement, "agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model."

For teams building document-processing or meeting-intelligence pipelines, that flexibility removes a common architectural headache: the separate vision or transcription model in the middle of the stack, with its own latency, cost and failure modes.

Where Ember fits in Meta's lineup

Ember is the fourth model under the Muse brand, and the release cadence tells you how fast Meta is iterating:

• Muse Spark — mid-weight model, released in April
• Spark Glimmer — Spark's open-source twin, released in August
• Muse Agent — flagship consumer AI product, released in September to unexpected success
• Muse Wonder — frontier model, released in October
• Muse Ember — lightweight execution model, released this Monday

The Muse Agent product has kept moving since launch, with new features including Spaces and video chat rolling out earlier this month.

The practical question

n What does this change for CIOs and platform teams? If your agent workloads are volume-bound — routing tickets, parsing long documents, firing tool calls at scale — Ember gives you a frontier-lab model at open-source-adjacent pricing, with the Wonder-pairing pattern as the intended deployment shape.

Ten cents per million tokens turns per-task AI cost from a line item you negotiate into one you largely stop watching.