What is Muse Ember?
Muse Ember is the smallest and cheapest model in Meta's Muse family, released on 9 November 2026 as muse-ember-1.0 on the Meta Model API and Muse Code. It reads text, images, audio, video and PDFs, writes text, and has a 1,048,576-token context with up to 524,288 output tokens. Reasoning effort runs from minimal to max, with max the default.
Meta pitches it as the model that does the repeated work under a larger planner: tool calls, coding subtasks, routing, extraction and long-document parsing, with Muse Wonder or Muse Spark deciding what to do. It is the first model Meta trained using Muse Wonder Full. The weights are not published; Muse Spark Open and Muse Glimmer are the downloadable Muse models.
How much does Muse Ember cost?
| Plan | Price | What you get |
|---|---|---|
| Up to 200K input tokens | $0.10 / $0.50 per 1M tokens | Input / output; cached input $0.004 |
| Over 200K input tokens | $0.30 / $1.50 per 1M tokens | Input / output; cached input $0.012; every token in the request billed at this rate |
Up to 200K input tokens: $0.10 input, $0.004 cached input, $0.50 output per 1M tokens. Over 200K: $0.30 / $0.012 / $1.50, applied to the whole request. Web search $2.50 per 1,000 queries.
Compare this price with other models
Is Muse Ember worth it?
Muse Ember's case rests on cost per finished task rather than price per token. Its list price, $0.10 per million input tokens and $0.50 per million output, is the same as Claude Haiku 5.5's. But Artificial Analysis scores it 36 on its Intelligence Index at max effort, level with Haiku 5.5, for about $0.14 a task against Haiku's $0.41, because it uses about a third as many output tokens and its cached input costs $0.004. Meta's launch post puts its own Muse Spark 1.3 one point higher at $2.39 a task, roughly 17 times as much.
The agent results are where it earns its place. On AutomationBench-AA, which runs business workflows end to end, Meta reports a strict pass@1 score of 37.8 against 13.9 for Haiku 5.5; Artificial Analysis's own run gives Ember 68.1%, level with Muse Wonder. It also beats Haiku on Terminal-Bench 4.0 (40.9% against 32.8%) in Artificial Analysis's testing. It loses on FrontierCode 2.0 (27.2% against 34.9%), and Meta itself says to hand hard single-shot coding, knowledge-heavy questions and visual coding to Spark or Wonder. Several figures in Meta's launch post come from its own internal suite, including the Personal Superintelligence Index, and its cost chart is labelled as estimates on proposed pricing.
Knowledge is the weak spot. Its AA-Omniscience accuracy of 21.1% is the lowest among models of similar intelligence, largely because it declines 64% of questions. The upside is a hallucination rate of 17.4%, half of Haiku 5.5's, which suits a worker that should say it does not know rather than guess.
The pricing has one trap. Any request over 200,000 input tokens, cached tokens and earlier conversation turns included, bills every token at $0.30 and $1.50, three times the base rate. An agent that drags a long history along will cross that line without anyone noticing, so the 1M-token window is cheaper to have than to use.
Where it earns its place
A platform team at a 300-person logistics company
Run thousands of small tool-calling steps a day inside an order-handling agent
A larger model plans and Ember executes the steps; on Artificial Analysis it costs about a third as much per task as Haiku 5.5. At 1 billion input and 100 million output tokens a month, all under 200K per request, the bill is about $150 before caching.
An operations analyst at an insurance broker
Pull fields out of long policy PDFs and scanned attachments
Ember reads PDFs and images directly. Keep each request under 200,000 input tokens: above that the whole request bills at $0.30 and $1.50 per million.
A developer already using Muse Spark
Cut the cost of routine sub-agent calls without rewriting the integration
Ember takes the same request body as Muse Spark, so most code changes only the model ID. Leave hard one-shot coding and fact-heavy questions on Spark, as Meta recommends.
Before you commit
- Setup
- A Meta Model API key and a call to muse-ember-1.0 through the Responses, Chat Completions or Messages endpoint; it is also on every Muse Code plan.
- Integrations
- Tool calling, structured output, computer use, image understanding and search grounding; the same request body as Muse Spark.
- Data and privacy
- Standard mode only on the API, and Standard-mode content is not used to train Meta models. No Contributor mode for Ember. Muse Code subscriptions start in Contributor mode, which allows training, until switched off. Closed weights, so no self-hosting.
- Cost at scale
- 1 billion input and 100 million output tokens a month cost about $150 under 200K tokens per request, or about $450 if every request goes over 200K. Cached input at $0.004 cuts repeated context further. No batch price is listed.
- Limits
- 1,048,576-token context, 524,288 output tokens, 3,000 requests and 4,000,000 tokens a minute per team, 600 background submissions a minute. Weak on factual recall; Meta advises against it for hard single-shot coding and visual coding.
What’s good
- Intelligence Index 36 for about $0.14 a task, against $0.41 for Haiku 5.5
- 37.8 on AutomationBench-AA pass@1 against 13.9 for Haiku 5.5, on Meta's numbers
- Text, image, audio, video and PDF input with a 1M-token context
- API content is not used to train Meta models
What’s not
- Requests over 200K input tokens cost three times as much
- Declines 64% of AA-Omniscience questions
- Trails Haiku 5.5 on FrontierCode 2.0
- Closed weights and only two places to run it
Best for
Teams running high-volume agent subtasks, extraction and routing who already pay for a larger planner model.
Muse Ember alternatives
Muse Ember: common questions
- Is Muse Ember worth it?
- Muse Ember is the cheapest way to get Haiku-class agent work done, provided you do not need it to know things. We rate it 4.2 out of 5.
- How much does Muse Ember cost?
- Up to 200K input tokens: $0.10 input, $0.004 cached input, $0.50 output per 1M tokens. Over 200K: $0.30 / $0.012 / $1.50, applied to the whole request. Web search $2.50 per 1,000 queries.
- What are the downsides of Muse Ember?
- Requests over 200K input tokens cost three times as much; declines 64% of AA-Omniscience questions; trails Haiku 5.5 on FrontierCode 2.0; closed weights and only two places to run it.
- Who is Muse Ember best for?
- Teams running high-volume agent subtasks, extraction and routing who already pay for a larger planner model.
- How much does Muse Ember cost for a team or in production?
- 1 billion input and 100 million output tokens a month cost about $150 under 200K tokens per request, or about $450 if every request goes over 200K. Cached input at $0.004 cuts repeated context further. No batch price is listed.
- Is Muse Ember safe for company data?
- Standard mode only on the API, and Standard-mode content is not used to train Meta models. No Contributor mode for Ember. Muse Code subscriptions start in Contributor mode, which allows training, until switched off. Closed weights, so no self-hosting.
- What are the best Muse Ember alternatives?
- The closest alternatives we have reviewed are Claude Haiku 5.5, GPT-6 Luna, Muse Spark 1.3 and Gemini 3.8 Flash.
Answers checked Nov 11, 2026.
Make Muse Ember? Show this rating on your site
