Meta Launches Muse Ember, a Low-Cost AI Model for Agent Workflows
Meta announced Muse Ember on November 9, 2026, introducing a new AI model designed to handle repetitive tasks inside agentic workflows at a lower cost. The model targets use cases such as tool calling, coding, data extraction, and long-document analysis, with API pricing starting at $0.10 per million input tokens and $0.50 per million output tokens.
Today we’re introducing Muse Ember, our smallest and most efficient model, built for complex reasoning execution and everyday agent work.
Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. With a 1M-token context window… pic.twitter.com/qqcjZw6L8x
– AI at Meta (@AIatMeta) November 9, 2026
Rather than relying on a large, expensive model for every step of a project, Meta positions Muse Ember as a specialized option for smaller tasks that need to run hundreds or thousands of times. It integrates directly with larger models such as Muse Wonder and Muse Spark, which handle planning and reasoning while Ember executes subtasks at scale. The model also supports a context window of up to one million tokens and accepts text, images, audio, video, and PDF files.
What Is Meta Muse Ember?
Muse Ember belongs to Meta’s Muse family of AI models, developed by Meta Superintelligence Labs. Its official API model ID is muse-ember-1.0. It represents the smallest and most efficient model in the lineup, built for tasks with clear, repeatable steps.
This approach is particularly relevant to AI agents. Unlike a conventional chatbot that responds to a single prompt, an AI agent can complete a sequence of operations: reading a file, searching for information, calling a tool, checking the result, and then moving to the next step.
In a software project, for example, a more capable model creates a plan and breaks it into smaller tasks. Muse Ember then handles requests to inspect files, run tests, and fix specific errors. A business processing thousands of documents uses it to classify contracts and extract selected clauses or data fields.
The underlying idea is to assign work according to its complexity and cost, rather than using the most powerful model at every stage.
Muse Ember’s Coding and Agent Capabilities
Meta has focused on tasks that frequently appear in software development and business workflows. These include calling tools as part of multi-step operations, classifying requests and routing them to the appropriate system, extracting information from files and logs, and carrying out specific coding tasks.
For software development, Muse Ember can read project files, run terminal commands and tests, and make targeted changes based on the assigned task. Its intended role includes repetitive operations within an agent session, such as inspecting sections of code, locating files, and addressing clearly defined errors.
Meta has also emphasized parallel processing for large collections of files, images, and logs. Potential applications include sorting customer support tickets, extracting information from thousands of documents, classifying images by content, and searching company records for specific information.
Several companies involved in early testing, including Box, Cline, Oracle, Cloudflare, and You.com, have reported use cases involving contract review, document analysis, and coding tasks. Cloudflare reported that Ember was around 30% faster in the slower portion of response times compared with the model it tested. You.com reported a 2.2-second response time at the 95th percentile in its coding tests.
Image, Audio, and Video Support With a One-Million-Token Context Window
Muse Ember is not limited to text input. It also accepts images, audio, video, and PDF files, while producing text output. This multimodal support can help with workflows involving several types of information, such as reading screenshots, extracting data from charts, or analyzing meeting recordings alongside related documents.
The model has a context window of 1,048,576 tokens, while its maximum output length is 524,288 tokens. This allows it to process large software projects or extensive collections of documents within a single request, reducing the need to divide the material into smaller sections.
These two limits serve different purposes. The context window determines how much information can be included in a request, while the output limit defines the maximum length of the model’s response.
Muse Ember Benchmark Results
Meta published comparisons involving Claude Haiku 5.5 and GPT-6 Luna. Independent evaluations from Artificial Analysis and Vals AI also provide additional performance data. The results show strong performance in agent and coding tasks, alongside clear trade-offs in complex software engineering benchmarks.
Artificial Analysis Intelligence Index
Muse Ember scored 36 points on the Artificial Analysis Intelligence Index. This matches Claude Haiku 5.5 at its highest reasoning level and exceeds GPT-6 Luna, which scored 28 points in the published comparison.
The cost per task offers another way to assess the model. Artificial Analysis estimated the cost of a task at approximately $0.14 for Muse Ember, compared with around $0.41 for Claude Haiku 5.5 at a similar score. Based on the evaluation published on November 9, this places Ember among the lower-cost options in its category.
Coding and Agent Benchmarks
In coding evaluations, Muse Ember scored 40.9% on Terminal-Bench 4.0, compared with 39.2% for Claude Haiku 5.5 and 16.4% for GPT-6 Luna. It also achieved 46 points on the Coding Agent Index-AA, ahead of Haiku at 37 points and Luna at 31.
The results were different on FrontierCode 2.0-AA. Ember scored 27.2%, while Haiku reached 34.9% and Luna scored 25.2%. This benchmark evaluates models on practical software engineering tasks, including the quality of code changes and their suitability for integration into projects.
On Agents’ Last Exam, Ember achieved 61.6%, compared with 57% for Haiku and 57.1% for Luna. It also scored 69.5% on OSWorld 2.1, which evaluates how models interact with computer applications using a partial-credit scoring system.
Vals Index Results
Muse Ember received 60.48% on Vals Index v2.2, placing it 12th among 35 models at the time of evaluation. Its Legal Index score was 75.69%, ranking fourth among 36 models, while its Finance Index score reached 60.91%.
The model also scored 25% on HLAB-Vals-1.1, placing it in second position alongside another model. However, its results were lower on some other specialized evaluations, including 39.39% on Terminal-Bench 4.0 and 38.23% on Tax Agent Bench.
These results show that Muse Ember’s performance aligns with its specialized design: it performed well in data classification, agent workflows, and legal evaluations, but trailed in complex software engineering and broader knowledge-based tests.
Muse Ember API Pricing and Availability
Meta has made Muse Ember available through the Meta Model API and Muse Code. API usage follows a pay-as-you-go pricing model, with rates depending on the number of input tokens.
| Usage type | Up to 200K input tokens | Over 200K input tokens |
|---|---|---|
| Input tokens | $0.10 | $0.30 |
| Cached input tokens | $0.004 | $0.012 |
| Output tokens | $0.50 | $1.50 |
Prices are per million tokens. When a request exceeds 200,000 input tokens, the long-context rates apply to the entire request, not only to the tokens above that threshold.
Muse Ember supports the Responses API, Chat Completions, and Messages interfaces. Its API model ID is muse-ember-1.0.
The model is available in Standard Mode through the Model API, without a Contributor Mode version. According to Meta’s official pricing information, the company does not use API request content to train its future models in this mode.
Muse Ember vs. Muse Spark and Muse Wonder
The Muse family includes models designed for different roles, with the choice depending on the work involved: Muse Wonder focuses on the most demanding reasoning and planning tasks, while Muse Spark targets coding and agentic workflows. Muse Ember is designed for repetitive coding and agent tasks that need to run at scale.
This division allows a larger model to handle planning before delegating numerous smaller operations to Ember. Meta also stated that Muse Ember is the first model in this category trained using Muse Wonder Full.
Safety and Model Access
Meta evaluated Muse Ember under its framework for assessing risks associated with advanced AI models. The evaluations covered cybersecurity, biological and chemical risks, and loss-of-control concerns. The evaluation confirmed that the model remained below the high-risk thresholds in the categories tested.
The safety evaluations also examined how the model handles sensitive information and user instructions, as well as its ability to resist malicious instructions hidden in web pages, documents, and tool outputs. Meta also highlighted the use of human approval before actions with significant real-world impact.
Muse Ember is currently available through the Meta Model API and Muse Code. Meta has not released the model weights for open download as part of the announcement.
Why Meta Built Muse Ember
With Muse Ember, Meta is targeting a major cost factor in agentic AI: the large number of repetitive operations involved in completing a task. Businesses do not always need a large model to analyze every file or classify every request. In many cases, a smaller model can handle a clearly defined operation at a lower cost.
This establishes Muse Ember as a practical, cost-effective solution for workflows that involve high volumes of repeated tasks, while larger models remain responsible for planning and more complex decisions. Meta’s approach puts the focus on the cost of completing an entire workflow, rather than performance on a single benchmark.
