Muse Ember
Muse Ember is Meta's smallest and most efficient model, released on Nov 9, 2026 for the repetitive work agents do: tool calls, coding subtasks, routing and extraction. It costs US$ 0.10 per million input tokens and US$ 0.50 per million output, the same base price as Claude Haiku 5.5 and GPT-6 Luna, with a cache 2.5 times cheaper. In Meta's table it leads both on 7 of 9 tests, ties on one and loses FrontierCode 2.0-AA to Haiku 5.5. Above 200 thousand input tokens, the whole request costs three times as much.
Muse Ember in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Meta (Meta Superintelligence Labs)
- API identifier
muse-ember-1.0- Release
- Nov 9, 2026
- Input
- US$ 0.10 per million tokens (up to 200 thousand input)
- Output
- US$ 0.50 per million tokens (up to 200 thousand input)
- Cache read
- US$ 0.004 per million (up to 200 thousand input)
- Above 200 thousand tokens
- US$ 0.30, US$ 0.012 and US$ 1.50, billed on the whole request
- Batch
- not published
- Context window
- 1 million tokens
- Max output
- 512 thousand tokens
- Input types
- text, image, audio and video
- Reliable knowledge
- not published
- Limits
- 3,000 requests and 4 million tokens per minute, per team
- Where it runs
- Meta Model API (Standard mode only), Muse Code, Google Cloud and Oracle Cloud
Where Muse Ember does well, where it falls behind and who it is for
Where it does well
- Process automation: 37.8 on AutomationBench-AA, against 13.9 for Claude Haiku 5.5 and 20.2 for GPT-6 Luna in Meta's table.
- Personal and office work: 46.8 on the Personal Superintelligence Index, Meta's internal index, against 36.1 for Haiku 5.5 and 37.0 for Luna.
- Cheap cache: rereading costs US$ 0.004 per million, against US$ 0.01 for Haiku 5.5 and GPT-6 Luna.
- Long answers: up to 512 thousand output tokens per call, against 128 thousand for Haiku 5.5 and Luna.
- The 100 thousand to 200 thousand token range: it stays at US$ 0.10 input, while Haiku 5.5 already charges US$ 0.50 above 100 thousand.
Where it falls behind
- Coding in real repositories: 27.2 on FrontierCode 2.0-AA, against 34.9 for Haiku 5.5.
- Very long context: above 200 thousand input tokens, the whole request moves to US$ 0.30 and US$ 1.50, more than GPT-6 Luna (US$ 0.20 and US$ 0.75 above 272 thousand).
- Thin margins in terminal and computer use: 40.9 against 39.2 for Haiku 5.5 on Terminal-Bench 4.0, with different execution environments, and 69.5 against 68.9 on OSWorld 2.1.
- The internal index changed: the Personal Superintelligence Index has eight evaluations for Ember and had seven for Muse Wonder, so the score does not compare across launches.
- Incomplete documentation: Meta has not published a batch price or a knowledge cutoff date.
Who it makes sense for
- A subagent that runs hundreds of short steps in a session (read a file, run a test, fix a snippet), directed by a larger model.
- Automation workflows across apps, where AutomationBench-AA shows the widest gap over its rivals.
- Reading many documents, images and recordings with a fixed cached instruction, at US$ 0.004 per million on rereads.
- High-volume extraction and routing, with requests under 200 thousand input tokens.
When to pick another model
- Code changes that have to be accepted by maintainers of real repositories: Haiku 5.5 scored higher on FrontierCode 2.0-AA.
- Requests that send more than 200 thousand input tokens at once: the price triples on the whole request.
- Hard reasoning problems: Meta itself points to Muse Wonder for planning and Ember for executing the volume.
- Using it through Muse Code with sensitive data without checking the settings: there, subscriptions come in the mode that uses content for training.
What Meta published at launch
Muse Ember came out on Nov 9, 2026 as the smallest model in the Muse family, below Muse Spark and Muse Wonder. Meta says it is the first model trained using Muse Wonder Full and that it was built for the work agents repeat all day. On the Meta Model API it runs in Standard mode only, where the content you send is not used to train Meta models.
The table below is the one from the announcement, with all three models at max effort. Two rows are left out: the Artificial Analysis index and the Coding Agent Index-AA, composite leaderboards the AI Benchmark does not reproduce. Meta's own model page in the documentation leaves those two rows blank.
The rival figures come from runs by Meta or third parties, on a best-effort basis, as the evaluation report notes. On Terminal-Bench 4.0, each model ran in a different environment: Ember in mini-swe-agent, Haiku 5.5 from the figure Anthropic published in Claude Code and GPT-6 Luna through Codex.
| Muse Ember Meta | Haiku 5.5 Anthropic | GPT-6 Luna OpenAI | |
|---|---|---|---|
| AutomationBench-AA business workflows |
37.8 | 13.9 | 20.2 |
| Agents' Last Exam agent tasks, Linux subset |
61.6 | 57.0 | 57.1 |
| Personal Superintelligence Index Meta internal index |
46.8 | 36.1 | 37.0 |
| Terminal-Bench 4.0 terminal coding |
40.9 | 39.2 | 16.4 |
| FrontierCode 2.0-AA code accepted by maintainers |
27.2 | 34.9 | 25.2 |
| GDP.pdf-AA work over long documents |
22.8 | 20.8 | 22.8 |
| OSWorld 2.1 computer use, partial score |
69.5 | 68.9 | 50.5 |
| OSWorld 2.1 computer use, full task |
33.8 | 27.8 | 17.0 |
| PerceptionBench visual perception, with tools |
80.6 | 78.6 | 62.8 |
Table from the Muse Ember announcement, with all three models at max effort. The dot marks the best in each row; on GDP.pdf-AA, Ember and GPT-6 Luna tie. The Artificial Analysis index and the Coding Agent Index-AA are left out.
Same price per token, different bills
Muse Ember, Claude Haiku 5.5 and GPT-6 Luna charge the same US$ 0.10 input and US$ 0.50 output per million tokens on short requests. The difference is at the edges: Ember's cache costs 2.5 times less, and each maker starts charging more at a different input size.
Haiku 5.5 changes tier above 100 thousand tokens, Ember above 200 thousand and Luna above 272 thousand. On Ember, a request past the limit is billed entirely at the higher tier, including cached tokens and the history of a stored conversation. Anyone sending large documents needs to measure where each request lands.
Meta claims Ember delivers more results per dollar than Haiku 5.5 and GPT-6 Luna at any reasoning setting. That calculation uses the Artificial Analysis index, which the AI Benchmark does not reproduce; it is covered in the article about the model, cited as Meta's figure.
| Muse Ember Meta | Haiku 5.5 Anthropic | GPT-6 Luna OpenAI | |
|---|---|---|---|
| Input | US$ 0.10 | US$ 0.10 | US$ 0.10 |
| Output | US$ 0.50 | US$ 0.50 | US$ 0.50 |
| Cache read | US$ 0.004 | US$ 0.01 | US$ 0.01 |
| Higher tier starts above | 200 thousand | 100 thousand | 272 thousand |
| Input and output, higher tier | US$ 0.30 and 1.50 | US$ 0.50 and 2.50 | US$ 0.20 and 0.75 |
| Max output | 512 thousand | 128 thousand | 128 thousand |
Meta Model API (Standard mode), Anthropic and OpenAI pricing pages, checked on Nov 9, 2026. Sizes in input tokens.
How Meta positions Ember in the Muse family
The family has three tiers: Muse Wonder for the hardest problems, Muse Spark for coding and agent workflows, and Muse Ember for repetitive tasks at scale. The documentation says Ember takes the same requests as Muse Spark at a lower price per token: US$ 0.10 and US$ 0.50, against US$ 1.25 and US$ 4.25 for Spark 1.3.
The suggested setup is a team: the larger model plans, and Ember executes the volume. In Meta's safety evaluation, Ember stayed below the high-risk thresholds for chemistry, biology, cybersecurity and loss of control.
Examples of using Muse Ember
A larger agent plans the change and dispatches file reading, test runs and targeted fixes to Ember. Cline describes exactly this use, and You.com measured a p95 of 2.2 seconds on the coding tasks it tested.
Box reports Ember scoring a contract clause by clause against the company's negotiation playbook. With the playbook cached, each reread costs US$ 0.004 per million tokens.
Pulling records, calling tools in sequence and updating systems is what AutomationBench-AA measures, across 657 workflows: Ember scored 37.8, against 13.9 for Haiku 5.5.
Where Muse Ember stands out and where it falls behind, by the numbers
Muse Ember has published results on 2 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Feb 12, 2026 and Feb 7, 2027 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
It is not in the bottom half of its generation on any test with 6 or more models compared.
See all 2 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| SimpleBench | 72.7% | #20 / 44 | #21 / 90 | 94.9% Claude Fable 5.2 |
| Apex Agents | 64.3% | #15 / 57 (tied with 1) | #15 / 60 | 84.2% Gemini 4 Carbon |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
Other models from Meta AI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Muse Ember | Nov 9, 2026 | no ECI | n/a | n/a | |
| Muse Spark Open | Oct 15, 2026 | 155.42 | n/a | n/a | Open weights (restricted use) |
| Muse Wonder | Oct 12, 2026 | 161.68 | 1,537 | n/a | API access |
| Muse Spark 1.3 | Sep 2, 2026 | 158.24 | 1,529 | 81.6 | API access |
| Muse Spark 1.2 | Aug 5, 2026 | 155.04 | 1,517 | 78.0 | API access |
| Muse Spark 1.1 | Jul 9, 2026 | 154.87 | 1,517 | 75.3 | API access |
| Muse Spark | Apr 8, 2026 | 153.14 | 1,507 | n/a | API access |
| Llama 4 Maverick | Apr 6, 2025 | 132.08 | n/a | n/a | Open weights (restricted use) |
| Llama 4 Scout | Apr 5, 2025 | 127.68 | n/a | n/a | Open weights (restricted use) |
Have you used Muse Ember? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Muse Ember
How much does Muse Ember cost?
On the Meta Model API, US$ 0.10 per million input tokens, US$ 0.004 for cache and US$ 0.50 for output, on requests of up to 200 thousand input tokens. Above that, the whole request costs US$ 0.30, US$ 0.012 and US$ 1.50. Web search costs US$ 2.50 per thousand searches.
Is Muse Ember better than Claude Haiku 5.5?
In Meta's table, Ember leads on 7 of 9 tests, ties one with GPT-6 Luna and loses FrontierCode 2.0-AA to Haiku 5.5 (27.2 against 34.9). The rival runs were best-effort, and on Terminal-Bench each model used a different environment. Independent scores, when they exist, are at the top of this page.
What is the difference between Muse Ember and Muse Wonder?
Wonder is Meta's model for the hardest problems and costs US$ 4 and US$ 20 per million tokens. Ember is the smallest and cheapest, at US$ 0.10 and US$ 0.50, built to execute volume. Meta suggests using them together: Wonder plans and Ember executes.
Does Muse Ember use my data to train models?
On the Meta Model API, Ember runs in Standard mode only, where content is not used for training. In Muse Code, subscriptions default to Contributor mode, which uses content for training, and it can be turned off in the settings.
Where can I use Muse Ember?
On the Meta Model API (identifier muse-ember-1.0, with Responses API support), in Muse Code, on Google Cloud and on Oracle Cloud. The limit is 3,000 requests and 4 million tokens per minute, per team.
What is Muse Ember best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Muse Ember are: Apex Agents (#15 of 57). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Meta, Introducing Our Smallest, Most Efficient Model: Muse Ember
- Meta, Muse Ember evaluation methodology (PDF)
- Meta, Muse Ember model page
- Meta Model API, pricing and rate limits
- Anthropic, pricing (Claude Haiku 5.5)
- OpenAI, API pricing (GPT-6 Luna)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
Read also: Muse Ember: what it is, pricing, pros and cons
← See the full ranking