# Meta launches Muse Ember, a cheap 1‑million‑token model for AI agents

> Meta’s new Muse Ember handles tool calls, coding subtasks and large document parsing at about a third the cost of similar models.

Oossa · 2026-11-10 · https://oossa.com/en/meta-launches-muse-ember-a-cheap-1-million-token-model-for-ai-agents

Meta’s Superintelligence Lab announced Muse Ember on Monday, November 9, 2026. The model is the smallest in the Muse family and is built for high‑volume agent work such as tool calls, coding subtasks, routing and parsing long documents. Meta says Ember can read text, images, audio and video and works with a context window of 1,048,576 tokens, producing up to 524,288 tokens in response.

The company positions Ember as the workhorse that pairs with larger models like Muse Spark or Muse Wonder. In an agent workflow, the bigger model plans and reasons while Ember handles repeated, low‑cost tasks across thousands of items. This design aims to cut per‑task costs dramatically.

## How Ember compares to rivals

Meta evaluated Ember on the Artificial Analysis Index, a benchmark that mixes coding, finance and business automation tests. According to Meta, Ember matched Anthropic’s Claude Haiku 5.5 on overall score while costing roughly a third as much per task. It also beat OpenAI’s Luna model on the same benchmark and led by 17 % on the AutomationBench‑AA suite of 657 cloud‑app tasks.

Pricing reflects the cost focus: 10 cents per million input tokens and 50 cents per million output tokens, with a cache that can drop input costs to 0.4 cents per million tokens. Long‑context requests over 200 K input tokens incur higher rates, as noted in the Model API documentation.

## Where Ember fits today

Muse Ember is available now on Meta’s Model API and through Muse Code under any pay‑as‑you‑go or subscription plan. It runs only in Standard mode, meaning users get the maximum hardware allocation for each request. The model accepts the same request format as Muse Spark, so switching models is a simple ID change in the API call.

Developers can use Ember for bounded work that repeats many times—like classifying support tickets, captioning large image sets, extracting fields from PDFs, or running terminal commands inside a larger workflow.

## The facts

- Muse Ember was announced on Monday, November 9, 2026 ("Meta Superintelligence Lab released a new AI model on Monday called Muse Ember").
- The model has a context window of 1,048,576 tokens and can output up to 524,288 tokens.
- Pricing is 10 cents per million input tokens and 50 cents per million output tokens; a cache can lower input cost to 0.4 cents per million tokens.
- Meta’s internal benchmark shows Ember matching Claude Haiku 5.5 at about a third of the price per task and leading by 17 % on the AutomationBench‑AA suite.
- Ember supports text, image, video, audio and PDF inputs and runs on Meta’s Model API in Standard mode only.

## Why it matters

For businesses that run large numbers of repetitive AI tasks—like customer‑support ticket routing or bulk document processing—Muse Ember offers a cheaper way to scale without sacrificing the ability to handle mixed media. The lower price per token can reduce operating costs, though independent third‑party validation of the benchmark claims is still pending.

## Sources & references

1. [Meta Releases Muse Ember, a 1M-Context Model for High-Volume Agent Work](https://www.marktechpost.com/2026/11/09/meta-muse-ember-small-agent-model/) – MarkTechPost, 2026-11-10
2. [Meta](https://dev.meta.ai/docs/muse-ember?project_id=1082575054202150&team_id=2306091636825706) – Meta
3. [Meta introduces Muse Ember to power cost-efficient AI agents](https://siliconangle.com/2026/11/09/meta-introduces-muse-ember-to-power-cost-efficient-ai-agents/) – SiliconANGLE, 2026-11-10
4. [Introducing Our Smallest, Most Efficient Model: Muse Ember](https://simonwillison.net/2026/Nov/10/muse-ember/) – Simon Willison, 2026-11-10
5. [Meta releases Muse Ember, a new lightweight model for agent execution](https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/) – TechCrunch, 2026-11-09

Last updated: 2026-11-10