Muse Ember review: cheap enough for the busy work, and Meta's own chart shows where to stop
Meta's smallest Muse model costs ten cents per million input tokens and wins most of the benchmarks Meta chose to publish. The most useful line in the launch is the one it loses, and the score it wins with still means most business workflows fail.

Muse Ember, released by Meta on November 9, 2026, is a strong candidate for the high-volume steps inside an AI agent: extraction, routing, tool calls and small coding fixes, at $0.10 per million input tokens and $0.50 per million output tokens with a 1M-token context. Meta's own chart shows it trailing Claude Haiku 5.5 on mergeable engineering work, and its best business-workflow score is 37.8% strict pass. Use it for steps you can check automatically, and keep a stronger model or a person on the rest.
What Meta released on November 9
Muse Ember is the smallest model in Meta's Muse family, and it's aimed squarely at the work agents repeat all day: tool calls, short coding tasks, routing, extraction and reading long documents. Meta describes the pairing plainly. "Muse Wonder plans, Muse Ember handles the volume."
The numbers that matter for a buyer:
- Price: $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, for prompts up to 200,000 tokens. Longer prompts cost three times as much.
- Size of work: a 1-million-token context window and up to 512,000 output tokens.
- Access: model id
muse-ember-1.0on Meta's Model API and Muse Code, with support for the Responses API. Meta says the Muse models are also available on Google Cloud and Oracle Cloud. - Inputs: text, images, audio and video, so an agent can read a scanned invoice or a screenshot without a separate model.
Meta also says it ran its usual safety evaluations, including how the model handles instructions hidden in web pages or documents and whether it seeks human input before high-impact actions.
Who it is for
Teams that already run an AI agent, or several AI steps inside an ordinary application, and are watching the bill grow with volume. The obvious candidates are support desks sorting and summarizing tickets, commerce teams cleaning and enriching product data, operations teams pulling fields out of invoices and shipping documents, and engineering teams whose coding agents send off many small subtasks per session.
It's less relevant if your AI use is a handful of chat sessions a day, where the price difference is pocket change, or if your hardest problem is judgment rather than volume. A cheaper model doesn't make a hard decision easier. It makes the easy steps around it cheaper to run, which frees budget for the stronger model, or the person, who handles the decision.
Meta's own positioning agrees. It places Ember below Muse Spark, which it recommends for coding and agentic workflows, and Muse Wonder, for the hardest problems.
A model that wins a business-workflow benchmark with 37.8% has still failed most of the workflows.
What the chart says, including the line it loses
Meta's launch includes a long benchmark table against Claude Haiku 5.5 and GPT-6 Luna, mostly from Artificial Analysis and partly from Meta's own suite. Ember wins most rows. Two of them are worth reading closely.
On AutomationBench-AA, which tests end-to-end business workflows with a strict pass-or-fail score, Ember scores 37.8%, against 20.2% for GPT-6 Luna and 13.9% for Haiku 5.5. That's a large lead. It also means that, by Meta's own measure, the best small model available fails most of the business workflows it attempts.
A model that wins a business-workflow benchmark with 37.8% has still failed most of the workflows.
On FrontierCode 2.0-AA, which judges real engineering tasks on whether the result could be merged, Haiku 5.5 leads with 34.9% to Ember's 27.2%. Meta published that loss in its own table, which deserves credit. It's also the most practical row in the chart: if the job is producing code a reviewer will accept, the cheaper model isn't the better one.
One more caveat. Meta's cost-per-task chart notes that Ember's costs are "estimates at proposed pricing." The headline claim, matching Haiku's best score at about a third of the cost, is Meta's analysis, not an independent measurement.
Decision guide
Where Meta's chart puts Muse Ember
| Consideration | Muse Ember (Meta's figures) | Best comparison model (Meta's figures) |
|---|---|---|
| AutomationBench-AA, end-to-end business workflows | 37.8% strict pass | GPT-6 Luna 20.2%; Claude Haiku 5.5 13.9% |
| FrontierCode 2.0-AA, mergeable engineering tasks | 27.2% | Claude Haiku 5.5 34.9% |
| Terminal-Bench 4.0, agentic terminal coding | 40.9% | Claude Haiku 5.5 39.2% |
| List price per million tokens (prompts up to 200K) | $0.10 input, $0.50 output | Muse Spark 1.3: $1.25 input, $4.25 output |
Where it is strong
Price for volume. At ten cents per million input tokens, reading a long contract or a full product catalog is cheap enough to run on every document, not just the ones someone flags. That changes which jobs are worth automating at all.
Long inputs. A million tokens of context lets one call take in a large document set. For extraction and summarizing, fewer calls usually means fewer places to lose track.
Speed. Launch partners quoted in the release describe it as fast. One reports it ran up to 30% faster at the tail than the comparable model it used, and another reports a 95th-percentile latency of 2.2 seconds on its coding tasks. These are partner claims, but they fit the model's purpose.
Mixed media. Reading screenshots, charts and recordings natively means fewer moving parts in an agent that handles messy business inputs.
Where it falls short
Final answers on hard work. The FrontierCode gap and the AutomationBench ceiling both say the same thing. Ember is good at steps. It isn't a safe owner for a whole workflow that ends in a customer-facing action.
Vendor numbers only, for now. Every figure in the launch comes from Meta or its partners. Benchmarks also reward what they measure. Your returns process, your product data and your edge cases weren't in the test set.
The long-prompt price. Prompts over 200,000 tokens cost $0.30 per million input tokens and $1.50 per million output tokens. That's still low, but if your plan depends on putting whole document sets into every call, price it at the higher rate.
Another vendor to manage. If you already run Anthropic or OpenAI models, adding Meta means another contract, another set of data terms and another API to monitor. For a small team, that overhead can outweigh the savings on a modest workload.
Our earlier note on routing cheap work to a small model still applies. The model changed. The rule didn't: send a step to the cheap model when you can check its output automatically.
How the work moves
Routing one support ticket through two models
- Read and sort
Ember reads the ticket and attachments, extracts order number and issue type.
- Look up
Ember calls the order and shipping tools and summarizes what it found.
- Decide
A stronger model, or a person, decides on refunds and exceptions.
- Check
Code checks every Ember field against the source before anything is sent.
Plain verdict
Muse Ember is a credible new option for the busy, checkable parts of an AI agent: sorting, extracting, looking things up, making small, testable code changes. The price makes it worth a serious trial, and Meta's own chart is honest about where it stops.
Before switching anything, run your own cases through it. Take a few hundred real examples from the step you want to move, score Ember against the model you use now, and decide on a pass mark in advance. Our note on setting an acceptance gate before a model takes the job walks through how. Keep a stronger model, or a person, on the decisions that cost money when they're wrong.
If you're building an agent and want help deciding which steps a cheap model can own, that's our AI applications and agents work. Talk to us.
Related work: Spin:Market
We build alongside the people doing the work.
Bring us the problem you’re thinking about.