/ai-search658 words

Meta's Muse Ember Undercuts Rival AI Agents at 10 Cents Per Million Tokens

Meta's Muse Ember, released Monday at 10 cents per million input tokens, targets high-volume agent work like tool calls, extraction and document parsing.

Meta releases Muse Ember, a new lightweight model for agent execution
Meta releases Muse Ember, a new lightweight model for agent executioncookieone / pixabay
  • Meta released Muse Ember on Monday at 10 cents per million input tokens.
  • Meta cited Artificial Analysis's composite benchmark to claim Ember is cheaper and smarter than Anthropic Haiku and OpenAI Luna.
  • Ember is the fourth Muse model, after Spark (April), Glimmer (August) and Wonder (October).
  • Ember handles text, images and audio inputs natively, per Meta's release.
  • The Muse Agent consumer product launched in September; Spaces and video chat shipped earlier this month.

Meta released Muse Ember on Monday, its lightest and most cost-efficient AI model to date, priced at 10 cents per million input tokens (a token is roughly three-quarters of a word, the unit AI models bill on). At that price, Ember undercuts frontier-class competitors Anthropic Haiku and OpenAI Luna while staying cost-competitive with open-source rivals DeepSeek V4.1 Flash and GLM 5.3 Flash.

The launch matters for search and automation watchers because of what Ember is built to do. Meta Superintelligence Lab said the model was "designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." Those are the high-volume, low-glamour tasks that dominate AI agent operations — and the ones where per-task cost decides whether an agent fleet is economically viable.

What does the benchmark claim rest on?

Meta's pricing and performance claims are confirmed company statements from the release. The company used Artificial Analysis's composite benchmark score (an independent lab's aggregate rating comparing models on reasoning, coding and cost) to position Ember as cheaper and smarter than Haiku and Luna.

That framing is vendor-selected, it's worth stressing: Meta chose the benchmark and the comparison set. What isn't in the release is third-party verification or raw benchmark tables, so treat the "cheaper and smarter" line as a company claim until independent evaluations land.

How does Ember fit into an agent stack?

The model is designed to pair with a heavier sibling in an agentic harness (a system that lets AI models call tools and run multi-step tasks autonomously). Meta's own copy makes the split explicit: "Muse Wonder plans, Muse Ember handles the volume."

A customer-support deployment illustrates the pattern. A frontier model like Muse Wonder sets the rules for handling an issue; Ember then applies those rules across tens of thousands of cases. The economics are the point — a 10-cent-per-million-input-token model can absorb repetitive execution work that would be prohibitive on frontier pricing.

What can Ember actually process?

Trained for volume-heavy use, Ember handles text, images and audio natively. Meta said "agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model."

That multimodal breadth in a budget-tier model is the more quietly significant detail. Extraction and document parsing workloads — exactly the tasks publishers and e-commerce operations route through automation — have traditionally required separate vision or transcription models, each adding cost and latency.

Where does Ember sit in the Muse lineup?

Ember is the fourth model under Meta's Muse brand:

  • Muse Spark, the 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 execution model, released this week

The consumer-facing product, the Muse Agent, launched in September and met what the company described as unexpected success. Spaces and video chat features rolled out earlier this month.

What's the competitive picture?

The sub-frontier tier is where the agentic AI market is currently fighting over unit economics. DeepSeek V4.1 Flash and GLM 5.3 Flash have pushed open-source prices down; OpenAI's Luna and Anthropic's Haiku hold the proprietary mid-tier. Ember's entry at 10 cents per million input tokens, with multimodal input included, pressures both camps on the metric that scales: cost per completed task.

For businesses running automated support, document processing or content-routing pipelines, the practical effect is a narrowing gap between "cheap enough to test" and "cheap enough to run at full volume."

Watch next for independent benchmark runs from Artificial Analysis and other evaluators, plus early adoption signals around the Muse Agent features shipped this month — those will show whether Ember's cost claims hold up outside Meta's own comparisons.

via techcrunch.com (Source)

More from James Calloway

James Calloway

Show full bio

Market editor covering money, funding and market shifts around Search rankings and algorithm updates, generative engine optimization at SERP Journal.

220 articles