Meta prices Muse Ember at 10 cents to undercut OpenAI and Anthropic

By Tom Whitfield

Meta will charge 10 cents per million input tokens for Muse Ember, its smallest and cheapest model to date, placing it well below proprietary small models such as Anthropic's Haiku and OpenAI's Luna. The release from Meta Superintelligence Labs targets the least glamorous side of AI agents: the repetitive background work, not one-off reasoning demos.

Ember handles tool calls, coding subtasks, request routing, data extraction and parsing long documents. That is work agents perform thousands of times a day, and it becomes expensive at scale. The 10-cent input price also puts Meta in the same range as low-cost open models including DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta has not emphasized output pricing in the same way — a gap that matters for agent workloads generating large volumes of text.

Can Ember's benchmark claims be trusted?

Meta claims Ember is both cheaper and smarter than rival lightweight models, citing Artificial Analysis's composite intelligence benchmark. Those comparisons come from the company's own announcement. Composite scores do not always reflect performance on specific real-world tasks. Independent testing on actual agent workflows will be the more meaningful measure.

How does Ember fit the Muse family?

Its role becomes clearer when paired with Muse Wonder, Meta's frontier model released in October. The larger model plans the work and sets the rules; Ember handles execution. A customer support system could use Wonder to decide how a complaint should be handled, then let Ember apply that approach across tens of thousands of tickets.

This planner-and-worker pattern is spreading across the industry because sending every step to a frontier model gets expensive quickly. Ember gives Meta a cheap worker unit for that architecture.

What inputs does Ember accept?

Ember takes more than text. Meta trained it to work with images and audio, so agents can process screenshots, charts and scanned documents, or work directly from voice memos, meeting recordings and video without switching models. For high-volume automation, avoiding that extra handoff saves both time and money.

A rapid build-out, completed

The release caps a fast expansion of the Muse family:

• Mid-sized Spark launched in April
• Glimmer, its open-source counterpart, arrived in August
• Consumer-focused Muse Agent launched in September and quickly became one of the most popular apps in the US
• Frontier model Wonder followed in October
• Muse Agent gained Spaces and video chat features this month

Meta has now covered each part of the AI stack, with Ember filling the high-volume end. The real test is whether developers trust Meta's benchmark claims enough to move their large-scale workloads onto the new model.