Skip to content
Theorem Weekly TW markTWTheoremWeekly

The evidence behind the week in machine learning.

Model updates

Does Meta Ember cut agent costs to 10 cents per million tokens?

Meta’s Superintelligence Lab released a lightweight model designed for high-volume agent tasks at a fraction of frontier costs.

Key points

  • Meta released Muse Ember, its lightest model yet, at 10 cents per million input tokens.
  • The model is designed for repetitive agent tasks like tool calls and document parsing.
  • Meta claims Ember is cheaper and smarter than competitors like Haiku and Luna.
Does Meta Ember cut agent costs to 10 cents per million tokens?
Depth
1 outlet · newest source 161h ago

Meta’s Superintelligence Lab released Muse Ember on Monday, describing it as the lab’s most cost-efficient model to date. The release positions the new model for complex reasoning execution and repetitive agent work.Meta launched a cheap AI model meant to handle boring, repetitive tasks for software agents.Meta introduced a low-cost execution model optimized for tool use and multi-modal input in agentic loops.

What was released or claimed

Meta stated that Ember is designed for complex reasoning execution and work agents repeat all day, including tool calls, coding subtasks, routing, extraction and long document parsing. The company highlighted its cost-per-task using Artificial Analysis’s composite benchmark score.

Who it affects

The model targets developers building agentic systems that require processing large volumes of cases. It is intended to pair with larger models like Meta’s Muse Wonder in a hierarchical setup.

What is still unknown

The material does not specify independent third-party validation of the claimed cost-per-task advantages. It also does not detail the exact performance drop-off when compared to frontier models on complex reasoning tasks.

So what

Meta suggests a workflow where a heavier model establishes rules and Ember applies them across tens of thousands of cases. This division of labor aims to reduce costs for volume-heavy use cases like customer support.

What the paper shows

The research tests Ember’s ability to handle multi-modal inputs including text, images and audio without separate models. Reported results claim it is cost-competitive with open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash. Limitations are not explicitly detailed beyond its role as a lightweight execution model.

Evidence: reportedMeta releases Muse Ember, a new lightweight model for agent execution

aifleet-techcrunch

Why it matters

  • Researchers can evaluate the trade-offs between cost and capability in hierarchical agent architectures.
  • The multi-modal input support allows agents to process screenshots and voice memos directly.
  • Comparison against open-source models provides a benchmark for cost-efficiency in agent workflows.
  • The release follows Meta’s broader Muse brand strategy including Spark and Wonder models.

Topics: Meta, AI agents, cost optimization, multi-modal models, Muse Ember

Updated

Then / now / next

Then

Meta’s Muse lineup included Spark, Glimmer and Wonder models.

Now

Meta has added Ember as its lightest and most cost-efficient model.

Next

Researchers may observe how widely Ember is adopted in agentic harnesses.

Story timeline

  1. Then Meta’s Muse lineup included Spark, Glimmer and Wonder models.
  2. 09 Nov 2026 Meta releases Muse Ember, a new lightweight model for agent execution
  3. Now Meta has added Ember as its lightest and most cost-efficient model.
  4. Next Researchers may observe how widely Ember is adopted in agentic harnesses.

The strongest doubt

Counterpoint from the sources

Meta’s claims rely on its own framing of Artificial Analysis benchmarks; independent verification of the cost-per-task advantage is not provided in the release.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

How much does Meta Muse Ember cost per million input tokens?

Meta states the cost is 10 cents per million input tokens.

Which competitors does Meta compare Ember against?

Meta compares Ember to Anthropic’s Haiku, OpenAI’s Luna, DeepSeek V4.1 Flash and GLM 5.3 Flash.

When was Muse Ember released?

Meta released Muse Ember on Monday, following the October release of Muse Wonder.

Connected stories

This storyMeta Muse and OpenAI Dots Target Job and Dining AutomationMeta Muse and OpenAI Dots TargetCan Amazon's Text-Based AI Agent Match Your Shopping Preferences?Can Amazon's Text-Based AI AgentDoes Oracle’s Converged Database Solve AI Agent Context Gaps?Does Oracle’s Converged DatabaseDo 150 AI agents at Anything startup replace human colleagues?Do 150 AI agents at Anything sta

Reading receipt

  • Meta released Muse Ember, its lightest model yet, at 10 cents per million input tokens.
  • The model is designed for repetitive agent tasks like tool calls and document parsing.
  • Meta claims Ember is cheaper and smarter than competitors like Haiku and Luna.

Still open: How does Ember perform on complex reasoning tasks compared to frontier models?

Still open: What are the specific latency implications of using Ember in high-volume agent workflows?

Time spent: 0:00 · Depth: Full

Sources

  1. TechCrunch AI

Related stories

Background reading