What Is Muse Ember? Meta’s 1M-Token Agent Model

Meet Muse Ember, Meta’s 1M-Token Agent Model Scoring 36 at $0.14 per Task

Muse Ember is Meta’s small, proprietary multimodal reasoning model for repetitive agent work, with a 1M-token context window and API pricing that starts at $0.10 per million input tokens and $0.50 per million output tokens. Meta Superintelligence Labs released it on November 9, 2026 for tool calls, coding subtasks, routing, extraction, and long-document parsing. At maximum reasoning effort, it scores a rounded 36 on the Artificial Analysis Intelligence Index and costs about $0.14 per benchmark task, matching Claude Haiku 5.5’s rounded score at roughly one-third of its benchmark cost.

Key Takeaways

• Positioning: Muse Ember is Meta’s smallest and most cost-efficient Muse model, built to execute high-volume steps inside larger agent systems.
• Capabilities: It accepts text, images, audio, and video, produces text, supports tool calling, and has a 1,048,576-token context window.
• Pricing: Standard rates are $0.10 input, $0.004 cached input, and $0.50 output per million tokens. Rates triple for requests above 200,000 input tokens.
• Performance: Meta’s launch table reports 35.8 on the Artificial Analysis Intelligence Index at max effort, rounded to 36, with a $0.136 benchmark cost per task.
• Access: Developers can use model ID muse-ember-1.0 through the Meta Model API or select Ember in Muse Code. Google Cloud public-preview and OpenCode integrations followed.
• Licensing: Muse Ember is proprietary. Its weights are private, and Meta has announced no open-weight version.

What Is Muse Ember?

Muse Ember is the execution tier of Meta’s Muse model family. It handles bounded, repeatable work while a larger model such as Muse Wonder plans difficult tasks. Meta summarizes the division as “Muse Wonder plans, Muse Ember handles the volume.”

Ember is the fourth Muse model, following Spark, Spark’s open-weight sibling Glimmer, and Wonder, according to TechCrunch’s release report. Meta also identifies Ember as the first released model trained using Muse Wonder Full, although the training process and architecture have not been detailed publicly.

The first version is Muse Ember 1.0. Meta’s Alexandr Wang later confirmed that crouton-alpha was an early version and said pre-release testing processed more than 5 trillion tokens.

Muse Ember is not the consumer Muse assistant. It is a developer model for agent infrastructure, and Meta has not said whether the consumer assistant will adopt it.

Muse Ember at a Glance

Specification	Muse Ember
Developer	Meta Superintelligence Labs
First release	November 9, 2026
Model ID	muse-ember-1.0
Type	Small proprietary multimodal reasoning language model
Primary role	Agent execution, tool calls, routing, extraction, document parsing, and coding subtasks
Input modalities	Text, image, audio, and video
Output modality	Text
Context window	1,048,576 tokens, commonly rounded to 1M
Maximum output	524,288 tokens, commonly rounded to 512K
Reasoning	Configurable effort; Artificial Analysis tested xhigh and max configurations
Pricing through 200K input tokens	$0.10 input / $0.004 cached input / $0.50 output per 1M tokens
Pricing above 200K input tokens	$0.30 input / $0.012 cached input / $1.50 output per 1M tokens
Headline benchmark	36 rounded AA Intelligence Index score; approximately $0.14 per max-effort task
Official availability	Meta Model API and Muse Code
Additional availability	Google Cloud public preview, OpenCode, and other developer integrations
Parameter count	Not yet confirmed
License and weights	Proprietary; private weights, with no open-weight release announced

How Muse Ember Works and What Makes It Different

Wonder-Led Orchestration is the central architectural pattern. A frontier model decomposes a problem, while one or many Ember instances perform the repetitive steps. Meta’s demonstrated example was a Python-to-Rust codebase conversion using 16 parallel agents orchestrated by Muse Wonder, as described in a developer discussion of the launch.

Volume Execution defines the intended workload. Ember reads files, calls tools sequentially, runs tests, applies targeted fixes, routes requests, and extracts structured information. Its multimodal input support lets one model process screenshots, charts, scanned documents, voice memos, meeting recordings, and video without a separate perception model.

Cached-Input Economics explains why identical sticker prices can produce different bills. Ember and Claude Haiku 5.5 both start at $0.10 input and $0.50 output per million tokens. Artificial Analysis attributes Ember’s benchmark advantage partly to 60% cheaper cached input and roughly one-third as many output tokens during its evaluation. A public pricing analysis reaches the same distinction between token prices and total task cost.

Source: @LuminaBench

Muse Ember’s price advantage is primarily a token-efficiency advantage, not a lower sticker price for ordinary tokens.

Max-by-Default Reasoning matters during migration. A third-party review of Meta’s documentation reports effort settings from minimal to max, with max as the default. Teams swapping model IDs should set effort explicitly rather than accidentally running every tagging or routing call at maximum reasoning.

Artificial Analysis lists 455 output tokens per second, a 4.69-second time to first answer token, and 3.1 minutes per max-effort Intelligence Index task. Its xhigh configuration completed tasks in 2.6 minutes at $0.12 each. These are evaluation measurements, not guaranteed production latency.

Long-Prompt Repricing separates capacity from affordability. Ember supports approximately 1M tokens of context, but its base prices apply only through 200,000 input tokens. Above that threshold, published rates rise to $0.30 input and $1.50 output per million tokens. A large context window therefore does not imply that every large prompt receives the headline price.

What You Can Do With Muse Ember

• Fan out coding work. A planner can assign Ember individual file edits, test runs, targeted repairs, and migration units. This fits bounded implementation work better than ambiguous system design or difficult merge decisions.
• Process enterprise documents. Early partner testing included scoring contracts clause by clause against negotiation playbooks, assigning overall risk ratings, and screening candidates against detailed briefs. Long-document extraction can also feed downstream structured workflows.
• Run support and operations triage. A larger model can establish handling rules, then Ember can classify tickets, tag records, retrieve order details, and apply those rules across tens of thousands of cases.
• Build multimodal agent pipelines. Ember can read visual documents or recordings and return text for routing, extraction, or tool invocation. It does not generate images, audio, or video.
• Prototype inexpensive generated interfaces. In one early community test, an animated voxel pagoda cost $3.09 with Ember versus $24.96 with Haiku 5.5. That is a single creative workload, not evidence of a universal eightfold saving.

Muse Ember is best understood as an execution layer for agents, not a universal replacement for a frontier reasoning model.

How Muse Ember Compares

Meta’s launch evaluation used raw Artificial Analysis results at maximum reasoning effort. Costs below describe that benchmark workload, not flat per-task API fees.

Model, max effort	AA Intelligence Index	Cost per AA task	Terminal-Bench 4.0	FrontierCode 2.0-AA
Muse Ember	35.8, rounded to 36	$0.136	40.9%	27.2%
Claude Haiku 5.5	36.2, rounded to 36	$0.411	39.2%	34.9%
GPT-6 Luna	27.8, rounded to 28	$0.152	16.4%	25.2%

At maximum reasoning effort, Muse Ember matches Claude Haiku 5.5’s rounded 36-point intelligence score while reducing the featured benchmark bill from $0.411 to $0.136 per task.

Source: @TeksEdge

Ember also leads Haiku on AutomationBench-AA, 37.8% to 13.9%, but loses on FrontierCode. Against GPT-6 Luna, Ember combines a higher intelligence score with a slightly lower benchmark cost. The Artificial Analysis release page independently lists Ember’s rounded score, approximate $0.14 task cost, and speed measurements.

Meta’s own comparison places Ember one intelligence point behind Muse Spark 1.3 while reducing cost from roughly $2.39 to $0.14 per task, an approximately 17-fold difference. For the planner-class end of the market, a related comparison of Kimi K3.1 Pro and Claude Opus 5.5 examines the capabilities that distinguish larger reasoning models.

Availability: How to Access Muse Ember

The primary access route is the Meta Model API using muse-ember-1.0. Documentation summaries list Responses, Chat Completions, and Messages endpoints, with Messages supporting Anthropic-style requests. Muse Code provides a second official route for coding and agent workflows.

Google Cloud received Ember in public preview alongside Muse Wonder and Muse Spark 1.3. OpenCode announced same-day support, and other developer gateways added integrations. Meta’s own API remains the clearest source for authoritative pricing, limits, and account requirements.

Muse Code subscriptions default to Contributor Mode, which provides more plan usage in exchange for allowing Meta to use content to train future models. Users can disable that setting. Documentation summaries state that Contributor Mode is unavailable on the API.

Engineers seeking a separate comparison baseline can test Claude Haiku 5.5 on kie.ai. Muse Ember itself is not currently listed in that catalog.

Muse Ember is proprietary: developers can rent hosted inference, but they cannot download its weights.

What We Don’t Know Yet

Meta has not published Ember’s parameter count, detailed architecture, training dataset, or a full technical account of how Muse Wonder Full contributed to training. “Smallest” currently describes its position in the Muse family rather than a disclosed size.

Production evidence remains thin. A separate public evaluation recorded 60.48% accuracy and $0.153 per test on its own aggregate index, illustrating how workload design changes both rankings and costs. Early community testing also includes reports of fast, inexpensive execution with below-average output quality on private tasks.

Knowledge reliability needs closer scrutiny. One preliminary community analysis of Artificial Analysis’s knowledge test reported 21.1% correct answers and 64% refusals; those figures require broader verification.

Finally, Meta has not provided a public supported-country list in the reviewed materials. One early user reported blocks in 20 countries, which remains unconfirmed. The precise billing treatment of cached tokens and stored conversations around the 200K threshold also deserves explicit official clarification.

Frequently Asked Questions

What is Muse Ember used for?

Muse Ember is used for high-volume agent execution: tool calls, routing, extraction, long-document parsing, and bounded coding subtasks. Meta designed Muse Ember to work under a larger planner such as Muse Wonder, rather than to own every difficult reasoning step.

How much does Muse Ember cost?

Muse Ember costs $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. Above 200,000 input tokens, published pricing summaries list $0.30, $0.012, and $1.50 per million tokens respectively. Artificial Analysis puts Muse Ember’s max-setting benchmark cost near $0.14 per task, which is a workload result rather than a flat task fee.

Is Muse Ember open source?

No. Muse Ember is proprietary, its weights are private, and Meta has not announced an open-weight release. Muse Ember is available through hosted products and APIs rather than as a downloadable model.

What is Muse Ember's context window?

Muse Ember has a 1,048,576-token context window, commonly rounded to 1M tokens. Muse Ember supports up to 524,288 output tokens, but prompts above 200,000 input tokens move into the higher price tier.

Muse Ember vs Claude Haiku 5.5: which is better?

Muse Ember is cheaper on the featured Artificial Analysis workload, while Claude Haiku 5.5 is stronger on some harder engineering tasks. At max effort, Muse Ember scores 35.8 and costs $0.136 per task; Haiku 5.5 scores 36.2 and costs $0.411. Haiku 5.5 leads FrontierCode 2.0-AA 34.9 to 27.2, so the better choice depends on the workload.

Is Muse Ember multimodal?

Yes. Muse Ember accepts text, image, audio, and video inputs and produces text. Meta also built Muse Ember for sequential tool calling and agent workflows across apps.

How can developers access Muse Ember?

Developers can call Muse Ember through the Meta Model API with model ID muse-ember-1.0, or select Muse Ember in Muse Code. Muse Ember also reached Google Cloud in public preview and OpenCode; Muse Ember is not an open-weight download.

What to watch next: look for a technical report disclosing architecture and training details, independent production tests that reproduce the $0.14-per-task economics, and official clarification of long-prompt billing and regional restrictions. Any change to private-weight licensing or confirmation that the consumer Muse assistant uses Ember would also materially change its role.

Try related models on Kie.ai: Claude Haiku 5.5, GPT-6 Sol and Luna and DeepSeek-V4.1-Flash.