# Muse Ember by Meta: specs, price, benchmarks and when to pick it

> What Muse Ember, Meta's small low-cost model, is: price, 1M context, the benchmarks Meta published and when to pick it. Data as of November 15, 2026.

*Tags: ai, models, meta, subagents, closed-models*

---


> [!info] Quick Definition
> **Muse Ember** is the smallest and most efficient model in Meta's Muse family, released on November 9, 2026 for repeatable agent work: tool calls, coding subtasks, routing and data extraction. You use it through the API as `muse-ember-1.0`, with a 1M-token context window and text, image, video, audio and PDF input.


## What is Muse Ember and when was it released?

Meta introduced it on November 9, 2026 on its research blog. In the Muse family each model has a role: Muse Wonder for the hardest problems, Muse Spark for coding and agent workflows, and Muse Ember for repeatable coding and agent tasks at scale. Meta also says Ember is the first model trained with Muse Wonder Full.

The practical idea: Ember is the cheap, fast piece you put to work many times, not the model you ask the hardest thing. To understand that difference, read [[small-vs-large-models|small vs large models]].

## What is it good for?

Meta recommends it for high-volume work: subagent tasks, tool calls, routing and processing large collections of images, files and records. Because it accepts text, image, video, audio and PDF, it fits tasks such as tagging hundreds of screenshots, extracting fields from documents or classifying support messages. For the concept, see [[multimodal-models|multimodal models]].

Reasoning effort ranges from minimal to max, and the default is max. For simple tasks, try a lower level and compare cost and quality.

Meta states its limits: for hard one-shot coding problems, knowledge-heavy questions or visual coding, use Muse Spark or Muse Wonder.

## What benchmarks did Meta publish and how do you read them?

Meta publishes these figures in the Ember announcement, with all models at max reasoning. Some are Meta internal evaluations and others come from Artificial Analysis. As of November 15, 2026:

| Benchmark | Muse Ember score | Measured by |
|---|---|---|
| Terminal-Bench 4.0 | 40.9% | Not stated in the announcement (each model's best published harness) |
| OSWorld 2.1 (partial credit) | 69.5% | Meta (internal evaluation) |
| OSWorld 2.1 (binary) | 33.8% | Table in Meta's announcement |
| Agents' Last Exam | 61.6% | Meta (internal evaluation, Linux subset) |
| AutomationBench-AA | 37.8% | Artificial Analysis (Intelligence Index v5); reported by Meta |
| FrontierCode 2.0-AA | 27.2% | Artificial Analysis (Intelligence Index v5); reported by Meta |
| PerceptionBench (with tools) | 80.6% | Meta (internal evaluation) |

How to read them:

- **Check which variant is reported.** OSWorld 2.1 appears with two figures, one with partial credit (69.5%) and one binary (33.8%). They are not the same thing: the binary one gives a much lower figure.
- **Meta also publishes the comparison.** In the announcement table, Meta places Claude Haiku 5.5 and GPT-6 Luna next to Ember. On FrontierCode 2.0-AA, Meta itself shows Haiku 5.5 ahead (34.9% versus 27.2%). A lab publishing an unfavorable case is a good sign for reading the table, but the table is still the lab's own.
- **Against its bigger sibling.** Meta cites Artificial Analysis (Intelligence Index v5): 36 for Ember and 37 for Muse Spark 1.3, with a cost per index task of 0.14 dollars versus 2.39.
- **Do not build rankings across labs.** Each one uses its own harness and settings (on Terminal-Bench 4.0, Meta used each model's best published harness), so the same benchmark name can give very different figures depending on who runs it. To go deeper, read [[ai-benchmarks|AI benchmarks]].

## How much does it cost and what context does it have?

Data as of November 15, 2026, according to Meta's documentation and announcement:

| Item | Value |
|---|---|
| API ID | `muse-ember-1.0` |
| Release date | November 9, 2026 |
| Context window | 1M tokens (1,048,576); maximum output 512K (524,288) |
| Input / output | Text, image, video, audio and PDF / text only |
| Price up to 200K input | 0.10 USD input, 0.004 USD cached input, 0.50 USD output (per million tokens) |
| Price above 200K input | 0.30 USD input, 0.012 USD cached input, 1.50 USD output (per million tokens) |
| Limits per team | 3,000 requests per minute and 4,000,000 tokens per minute |
| Open or closed | Closed: offered through the API, not among Meta's open-weight models |
| Mode | Standard only on Meta's API |

One billing detail: when a request goes over 200K input tokens, the whole request is billed at the long-context rate, not just the tokens above the threshold. Cached tokens count toward that total.

Illustrative example of a calculation: a request with 50,000 input tokens and 5,000 output tokens costs 50,000 x 0.10 / 1,000,000 = 0.005 USD plus 5,000 x 0.50 / 1,000,000 = 0.0025 USD, about 0.0075 USD. To understand each piece, see [[tokens|tokens]], [[cost-per-token|cost per token]] and [[context-window|context window]].

## When should you pick it and when not?

**Pick it when:**

- You have thousands of items to classify or extract and cost per request decides.
- You need mixed input (screenshots, audio, PDF) in a single model.
- You are building a subagent workflow and want the repetitive piece to be cheap.

**Do not pick it when:**

- The problem is hard and one-shot; Meta recommends Spark or Wonder.
- You need open weights to run it yourself. See [[open-vs-closed-models|open vs closed models]].
- You want to compare with a small model from another lab: also test [[claude-haiku-5-5|Claude Haiku 5.5]] on your own data.

## Example

Illustrative example: a product designer exported 200 user comments, some with a screenshot, and needs to group them by theme for a prioritization session. She writes an instruction asking, for each comment, for a JSON with theme, severity from 1 to 3 and a short quote, and runs it with Muse Ember. With about 2,000 input tokens and 300 output tokens per comment, the batch comes to roughly 0.07 USD at the rates in the table.

What to try tomorrow: take 20 comments, classify them by hand and compare with the model's result. If they match on most, scale to the full batch; if not, adjust the instruction or raise the reasoning level.

## Common mistakes

- Using Ember for the hardest problem in the project instead of Spark or Wonder.
- Reading a single benchmark number without checking the variant (partial credit or binary).
- Assuming it has open weights because Meta has other open models.
- Leaving reasoning at max for trivial tasks without comparing cost.
- Crossing 200K input tokens without noticing that the whole request changes rate.
- Scaling to the whole batch without checking a sample by hand first.

## Sources

- [Meta: Muse Ember announcement](https://research.meta.ai/blog/muse-ember)
- [Meta: models documentation](https://dev.meta.ai/docs/models)
- [Meta: Muse Ember model page](https://dev.meta.ai/docs/muse-ember)
- [Meta: pricing and rate limits](https://dev.meta.ai/docs/pricing-rate-limits)


---

Source: https://www.fernandoux.com/en/wiki/ai/muse-ember/
