Updated November 10, 2026
What is Muse Ember?
Muse Ember is Meta's small, cost-efficient model for the repetitive work inside agent systems: tool calls, coding subtasks, routing, extraction and long-document processing. Its model ID is muse-ember-1.0, and Meta makes it available through Muse Code and the Meta Model API. The release lists a one-million-token context window and up to 512,000 output tokens. Ember accepts text, images, audio and video, making it relevant to workflows that combine documents, screenshots, recordings and structured actions. Its value proposition is useful capability at high volume, rather than replacing every larger model with the cheapest possible request.
Shortlist Ember for high-volume agent steps where you can validate outputs and escalate hard cases. Its $0.10 input and $0.50 output token rates are attractive, but cost per accepted result matters more than token price alone.
Official model and release documentation
Specification | Verified detail |
|---|---|
Provider | Meta |
Model ID | muse-ember-1.0 |
Access | Meta Model API and Muse Code |
Distribution | Hosted API; no weight release established by cited announcement |
Context window | 1M tokens |
Maximum output | Up to 512K tokens |
Input modalities | Text, images, audio and video |
API support | Responses API, according to Meta |
Positioning | Repeatable coding and agentic work at scale |
Standard input/output | $0.10 / $0.50 per million tokens |
Cached input | $0.004 per million tokens |
What changed and how to approach migration
Ember extends the Muse lineup rather than simply renaming Muse Wonder or Muse Spark. Meta says it is the first specialized model trained using Muse Wonder Full and positions Wonder for the hardest problems, Spark for coding and agentic workflows, and Ember for repeated tasks at scale. That suggests a practical division of labor: a stronger model can plan or resolve ambiguity while Ember handles bounded execution. This is a deployment pattern to test, not a guarantee that a mixed-model system will be cheaper once routing mistakes, retries and integration work are included.
Area | What changed |
|---|---|
Lineup role | Adds a low-cost execution tier to Muse |
Training relationship | Meta says it was trained using Muse Wonder Full |
Hardest tasks | Meta points to Muse Wonder |
Everyday agents | Meta also offers Muse Spark |
Integration implication | Evaluate routing and escalation, not only a model swap |
Capabilities and practical use cases
Ember's natural home is an agent that performs many small, inspectable steps. It can extract fields from a document, select an approved tool, make a targeted code change or classify an incoming request before a larger model reviews exceptions. Its multimodal input support also makes it a candidate for screenshot-based interaction and document workflows containing charts or recordings. The release describes computer use and long-document tasks, but native input support should not be confused with every tool being available in every product. Verify the exact endpoint, tool interface and supported file handling for the application you intend to build.
Workflow | Practical fit |
|---|---|
Routing | Choose the right downstream workflow from a clear schema |
Extraction | Turn source material into checkable structured fields |
Coding subtasks | Make narrow changes and run focused validation |
Tool sequencing | Complete bounded multi-step business workflows |
Multimodal review | Read screenshots, charts, recordings and video inputs |
Document processing | Work over large collections without losing source references |
Benchmarks and evidence
Meta's release reports an Artificial Analysis Intelligence Index score of 36 for Ember at max reasoning, alongside 36 for Haiku 5.5 and 28 for GPT-6 Luna in that table. It also reports 37.8 on AutomationBench-AA and 40.9 on Terminal-Bench 4.0. These figures are presented by Meta, and some comparisons use each model's best published harness rather than a single shared agent configuration. The reported approximately $0.14 cost per Artificial Analysis task is tied to that workload and list pricing; it is not a quote for processing an arbitrary customer task.
Evidence | Reported result | How to interpret it |
|---|---|---|
Artificial Analysis Index | 36 | Meta release; max-reasoning comparison |
AutomationBench-AA | 37.8 | Reported end-to-end business workflow score |
Agents' Last Exam | 61.6 | Meta-reported comparison |
Terminal-Bench 4.0 | 40.9 | Best published harness conditions matter |
Cost per AA task | About $0.14 | Benchmark-specific, not universal pricing |
Pricing and total cost
Meta lists $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. At those base rates, one million uncached input tokens plus 100,000 output tokens costs $0.15. Ten million input tokens and one million output tokens costs $1.50 before any other charges. These are arithmetic examples, not measured application costs. Include reasoning and generated-token usage, external tools, storage and retries when comparing providers. A cheap request that repeatedly produces invalid tool arguments can be more expensive operationally than a stronger model's first successful attempt.
Cost component | Published term or planning boundary |
|---|---|
Standard input | $0.10 / million tokens |
Cached input | $0.004 / million tokens |
Standard output | $0.50 / million tokens |
1M input + 100K output | $0.15 base token illustration |
10M input + 1M output | $1.50 base token illustration |
Benchmark cost claim | About $0.14 per AA task, workload-specific |
Application budget | Add tools, retries, storage and review time |
Limitations and deployment tradeoffs
Low token cost is most useful when the task has a clear acceptance test. Ambiguous planning, unusual edge cases and difficult technical reasoning may still need a stronger model. The release's favorable comparisons do not establish that Ember beats every effort setting, harness or language in your own application. A large context window can also hide costs if the system resends unnecessary material. Keep source provenance attached to extracted facts, validate tool arguments before execution and measure how often the workflow escalates. Multimodal understanding is a reason to pilot the model, not evidence that every media format or output modality is supported.
Consideration | What to check |
|---|---|
Ambiguous tasks | Escalate rather than forcing a cheap answer |
Invalid structured output | Validate fields and tool arguments |
Repeated context | Measure cache hits and avoid redundant material |
Vendor comparisons | Check effort and harness before claiming a win |
Multimodal handling | Verify actual endpoint and media limits |
How to evaluate Muse Ember for your workload
Start with a small, representative set of high-volume agent execution tasks, including ordinary cases, difficult edge cases and tasks with an objectively checkable answer. For this model, a useful pilot includes a routing-and-extraction queue with strict schemas and an escalation path. Keep the prompts, available tools, source material and acceptance criteria consistent across candidates. Record the exact model ID, reasoning configuration, date and any provider-specific access program so the comparison can be repeated later.
Measure completed-task quality rather than the fluency of the response. Ask a reviewer to identify unsupported claims, incorrect changes, missing evidence and avoidable manual repair. An agent that stops honestly when evidence is insufficient can be more useful than one that produces an impressive-looking but invalid artifact. Where tools are involved, track invalid arguments, failed calls and recovery behavior separately from the final answer. A single successful demonstration is a reason to continue testing, not proof of a dependable service level.
Pilot measure | What to record |
|---|---|
Task completion | Was the requested artifact actually completed and checked? |
Evidence quality | Can a reviewer trace decisive claims to sources or tests? |
Correction burden | How much expert time was needed after the model finished? |
Latency | Measure end-to-end time, not just first-token speed |
Total cost | Include tokens, tools, retries and human review |
Failure handling | Did it report uncertainty and recover without inventing success? |
Use a staged rollout only after the pilot meets the target you set in advance. Preserve a known baseline and a way to compare new requests with the previous configuration. If the model, tool interface, access policy or price changes, rerun the affected cases instead of assuming an old result still applies. The purpose of evaluation is a defensible choice for your actual work, not a claim that one model is universally best.
Who should use it, and what should you compare?
Ember fits product teams processing many similar requests and engineering teams able to define narrow agent subtasks. It can be especially interesting when most work is straightforward but a small share requires expert reasoning. A two-tier system should record both the cheap path and escalation path so savings are measured honestly. For complex architecture decisions or open-ended investigations, compare Muse Wonder or another frontier model. The useful question is not whether Ember is universally smarter than a competitor, but whether it meets your service level at a lower total operating cost.
Alternative | Why compare it | Decision boundary |
|---|---|---|
Muse Wonder 1.0 | Difficult planning and technical work | Use as an escalation candidate |
Muse Spark 1.3 | Coding and agentic workflows | Compare real workflow cost, not only rates |
Claude Haiku 5.5 | Latency-sensitive model alternative | Match effort and harness in your test |
GPT-6 Luna | Lightweight OpenAI option | Evaluate with identical schemas and checks |
Frequently asked questions
Is Muse Ember open-weight?
The cited release establishes hosted availability through Meta products and API, not a downloadable weight release. Do not infer open-weight licensing from Meta's broader Llama history or from the model's small-model positioning.
Can Ember work with images and recordings?
Meta says it natively understands text, images, audio and video. Confirm the supported input packaging, limits and tools in the actual endpoint you use; this does not establish image or audio generation output.
Should Ember replace Muse Wonder?
Not automatically. Meta positions Wonder for the hardest problems and Ember for repeated execution work. A routing pilot can show which tasks remain reliable on Ember and which need the stronger model.
Is $0.14 the cost of any agent task?
No. That figure is Meta's benchmark-specific cost per Artificial Analysis task at list pricing. Your task may use a different context length, number of steps, output budget, retry rate and set of paid tools.
Related model guides
Compare the full AI model directory
Sources and editorial method
Meta Muse Ember release and pricing
Specifications, access conditions and prices were checked against the linked provider sources on November 10, 2026. Provider-reported benchmarks are attributed and are not presented as independently reproduced tests. Missing specifications remain unverified rather than being inferred from a related model. Practical workflow advice is an evaluation recommendation, not a claim that we ran new hands-on trials. Recheck official terms before purchase or deployment.