# Meta's Muse Ember: a $0.10-per-million-token agent model with 1M context. What the benchmarks show, and where it trails Haiku

> Meta released Muse Ember, the smallest model in its Muse family, priced at $0.10 per million input tokens and $0.50 per million output tokens. Meta says it beats Claude Haiku 5.5 and GPT-6 Luna on most agent benchmarks. Its own chart shows Haiku ahead on real-world coding.

Published Nov 9, 2026 by TechMeasured Staff · AI · Canonical: https://techmeasured.com/ai/meta-muse-ember-10-cents-per-million-tokens-1m-context-agent-model-what-the-numbers-show

Meta released **Muse Ember** on November 9, calling it the smallest and most efficient model it has made. It's a cheap, fast model for the high-volume parts of agent work: tool calls, coding subtasks, routing, data extraction and long-document parsing. According to [Meta's announcement](https://research.meta.ai/blog/muse-ember), it costs $0.10 per million input tokens and $0.50 per million output tokens. It has a 1-million-token context window and accepts text, images, audio and video.

Why it matters: agents make many small, repetitive calls for every hard planning step, and Meta is pitching Ember for that bulk work. It's meant to sit underneath Muse Wonder, its October frontier model. [TechCrunch](https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/) reports Meta's framing: Wonder plans and sets the rules, Ember applies them at volume. It's also Meta's direct challenge to Anthropic's Claude Haiku and OpenAI's GPT-6 Luna in the cheap-model tier. Anthropic released Claude Fable 5.2 and Mythos 5.2 the same day.

## Price and specs

| | Muse Ember (`muse-ember-1.0`) |
|---|---|
| Input (prompts ≤200K tokens) | $0.10 per million tokens |
| Cached input | $0.004 per million |
| Output | $0.50 per million |
| Input / cached / output (prompts >200K) | $0.30 / $0.012 / $1.50 per million |
| Context window | 1M tokens, up to 512K output |
| Modalities in | Text, images, audio, video |
| Where | Meta Model API, Muse Code, Google Cloud, Oracle Cloud |

Meta's page doesn't mention open weights or a license, and every channel it lists is a hosted API. That sets it apart from Muse Glimmer, the open counterpart to Muse Spark that [TechCrunch](https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/) says came out in August.

## The benchmarks Meta published

All figures are Meta's, at maximum reasoning effort. Results labelled "-AA" come from Artificial Analysis evaluations that Meta cites.

| Benchmark | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| AutomationBench-AA | **37.8%** | 13.9% | 20.2% |
| Agents' Last Exam | **61.6%** | 57.0% | 57.1% |
| Terminal-Bench 4.0 | **40.9%** | 39.2% | 16.4% |
| Coding Agent Index-AA | **46** | 37 | 31 |
| FrontierCode 2.0-AA | 27.2% | **34.9%** | 25.2% |
| OSWorld 2.1 (partial / binary) | **69.5% / 33.8%** | 68.9% / 27.8% | 50.5% / 17.0% |
| PerceptionBench (with tools) | **80.6%** | 78.6% | 62.8% |
| Artificial Analysis Intelligence Index | 36 | 36 | 28 |

Two numbers stand out. On the Artificial Analysis Intelligence Index, Ember **ties** Haiku 5.5 rather than beating it. Meta's argument is about price: it says Ember gives more intelligence per dollar than Haiku, Luna and other flash-tier models at every reasoning setting. Second, Meta says an Artificial Analysis task costs about **$0.14** on Ember, compared with **$2.39** on its own Muse Spark 1.3, which scores just one point higher (37). If that holds up, Ember is roughly 17 times cheaper per task for nearly the same index score. That's the most useful figure for anyone running Spark-class workloads today.

The weak spot is coding. On FrontierCode 2.0-AA, which tests real-world coding, Haiku 5.5 leads by almost eight points. [The Decoder](https://the-decoder.com/meta-launches-muse-ember-a-small-ai-model-built-for-the-repetitive-work-agents-actually-do/) also flags this gap. It points out that most of the performance claims come from Meta's own testing. The independent figures are limited to the Artificial Analysis index, per-task cost and the FrontierCode result.

## Safety

Meta says Ember got the same core safety training as its larger models. It says it evaluated Ember under its Superintelligence Scaling Framework for chemical and biological, cyber and loss-of-control risks, and that the model stayed below the high-risk thresholds. Meta also says it tested Ember on ignoring instructions hidden in web pages and tool outputs, and on asking a human before high-impact actions. Both matter for a model built to run unattended in agent loops.

## Who should care

- **Teams paying flash-tier prices for agent sub-steps** (classification, routing, extraction, tool-call chains). Ember's input price and very cheap cached input are worth a side-by-side test on your own traces.
- **Coding-agent builders** should test before switching. Meta's own chart shows Haiku 5.5 ahead on real coding.
- **Anyone needing open weights or on-prem deployment**: Meta hasn't announced either for Ember. Meta's open option in the family is still Muse Glimmer.

What to watch: independent results from Artificial Analysis and other evaluators across the full benchmark set. Also watch whether Anthropic or OpenAI cut prices in this tier in response.


## Sources

1. [Meta AI — Introducing Our Smallest, Most Efficient Model: Muse Ember](https://research.meta.ai/blog/muse-ember)
2. [TechCrunch — 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/)
3. [The Decoder — Meta launches Muse Ember, a small AI model built for the repetitive work agents actually do](https://the-decoder.com/meta-launches-muse-ember-a-small-ai-model-built-for-the-repetitive-work-agents-actually-do/)
