# Meta Launches Muse Ember to Undercut Premium Small Models in Agent Workloads

> The fourth model in Meta's Muse family pairs a 1-million-token multimodal context window with per-task pricing that the company says matches Claude Haiku 5.5 at about a third of the cost and beats GPT-6 Luna.

*Published 2026-11-09 · By Marcus Vance*

Muse Ember is Meta's smallest and most efficient model, a multimodal model with a 1,048,576-token context window that Meta is positioning for the high-volume agent work of tool calls, coding subtasks, routing, extraction, and long document parsing.

Meta AI Research introduced Muse Ember in a launch post that calls the model 'our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day.' The model is available immediately as muse-ember-1.0 through the Meta Model API and Muse Code, making it the fourth release in the Muse family after Muse Spark, Muse Glimmer, and Muse Wonder.

Meta's positioning separates Ember from general-purpose flagships. The Meta Model API documentation describes it as 'the efficient execution model in the Muse family, built for high-volume, repeated work,' and says it runs the sub-agent tasks, tool calls, and routing inside larger agent workflows while processing large collections of images, files, and records.

## What is Muse Ember and where does it fit in Meta's model family?

The 'small' label refers to operational footprint, not context capacity. Muse Ember has a 1,048,576-token context window and supports up to 524,288 output tokens. Meta says the model handles complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. The model also has native multimodal support for text, images, video, audio, and PDF input.

Meta designed Ember to work as the executor inside a two-model workflow. The company says Ember is built to pair with larger models such as Muse Wonder or Muse Spark, where the larger model plans the overall task and Ember handles high-volume subtasks. That division of labor is aimed at teams that want a cheaper model for repeated operations without giving up the reasoning of a frontier-scale planner.

## What new capabilities does Muse Ember bring to agent workflows?

The feature set targets the unglamorous work that generates most of the volume in production agents. Meta's launch material lists five categories explicitly: tool calls, coding subtasks, routing, extraction, and long document parsing. Each of those tasks has been a common failure point for small models because they require following instructions precisely over many turns or across very long inputs.

- Tool calls: Ember runs tool calls inside larger agent workflows, a role Meta's Model API documentation assigns to the efficient execution model.
- Coding subtasks: Meta lists coding subtasks among the repeated work Ember is built to handle, alongside routing, extraction, and parsing.
- Routing: The model directs work to downstream models or functions, a function that scales with context and instruction-following accuracy.
- Extraction: The 1,048,576-token context window lets Ember pull structured information from long documents and large collections of files and records.
- Long document parsing: Native multimodal input for text, images, video, audio, and PDFs extends that parsing to documents that mix media and PDF formats.

The context window is the most consequential specification for enterprise deployment. A model that can accept 1,048,576 tokens in a single request can process entire contracts, regulatory filings, or technical manuals without the chunking strategies that older small models required. The output limit of 524,288 tokens is similarly large for a model in this tier, though the standard API pricing mode applies to requests up to 200,000 input tokens, according to Meta AI Research.

## How do Muse Ember's benchmark results and pricing compare with Claude Haiku 5.5 and GPT-6 Luna?

Meta published two head-to-head comparisons in its launch post. On Artificial Analysis benchmarks, the company says Muse Ember matches Claude Haiku 5.5's best score at about a third of the cost per task and beats every GPT-6 Luna setting. On AutomationBench-AA at maximum reasoning, the reported gap is wide: Muse Ember scores 37.8%, compared with 13.9% for Claude Haiku 5.5 and 20.2% for GPT-6 Luna.

*Reported comparisons for Muse Ember, Claude Haiku 5.5, and GPT-6 Luna*

| Metric | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
| --- | --- | --- | --- |
| AutomationBench-AA score at maximum reasoning | 37.8% | 13.9% | 20.2% |
| Relative cost per task on Artificial Analysis benchmarks | About one-third of Claude Haiku 5.5's cost | Baseline comparison | Beats every GPT-6 Luna setting |
| Token price, standard mode | $0.10 per million input; $0.50 per million output | Not provided in Meta's launch post | Not provided in Meta's launch post |

Token price is not the same as cost per task. A model that needs more output tokens or more retries can erase a cheap input price, which is why Meta and benchmark providers emphasize cost per completed task. The company's 'about a third of the cost' claim refers to per-task cost on Artificial Analysis benchmarks, not to the sticker price of each token.

The pricing itself is positioned below the premium small-model tiers that Anthropic and OpenAI sell for agent-heavy workloads. In standard mode, Muse Ember is priced at $0.10 per million input tokens and $0.50 per million output tokens, according to Meta AI Research. For teams running thousands of routing or extraction calls a day, that price changes the unit economics of delegating work to a subagent rather than a larger model.

## Why does the pricing model matter for high-volume agent tasks?

The economics matter because agent workloads are repetitive by design. A planner model may call a subagent dozens of times in a single session, and each call consumes tokens on both sides. Meta's reference architecture addresses that pattern directly: Muse Wonder or Muse Spark handles planning, while Muse Ember handles the high-volume subtasks. That pairing lets teams reserve expensive context for the planner and use the cheaper model for the repeated execution work.

Meta also frames the model around real enterprise content. Yashodha Bhavnani, VP of AI Products at Meta, cites document review and candidate screening as examples of what Ember can do inside an agent harness. The emphasis on contract scoring and playbook compliance points to procurement, legal, and HR use cases where accuracy and cost per review are measurable.

## What do developers and enterprise customers say about Muse Ember?

Meta's launch post includes a customer-facing endorsement from Yashodha Bhavnani, VP of AI Products, who describes the model's performance on the Box AI Complex Work Eval.

> On the Box AI Complex Work Eval, Muse Ember delivers impressive accuracy, speed and token efficiency. It can score a contract against a company's negotiation playbook clause by clause and land on the right overall risk rating, or screen candidates against a detailed brief and flag the borderline ones who still qualify as alternates. Muse Ember is a fast, efficient foundation for agents doing real work on enterprise content.Yashodha Bhavnani, VP of AI Products

Meta also published a developer reaction from Saoud Rizwan, founder of an unidentified company, who calls Ember a fast, low-cost model for subagent work. The quote emphasizes the two patterns that matter to developers: using Ember directly for fast, reliable inference, and running it inside a harness as the subagent to cut costs.

> It's incredible seeing Meta release a lightweight model for developers who want a fast, low-cost model they can rely on, and for teams cutting costs by running it as the subagent in their harness. Muse Ember is purpose-built for, and excels at, subagent work like coding and document analysis.Saoud Rizwan, Founder

The two quotes reflect the same division of labor. Enterprise buyers are evaluating Ember on accuracy, speed, and token efficiency, while developers are evaluating it as a drop-in executor for coding and document analysis. Both groups are responding to a common constraint: agent workflows generate far more subtask traffic than planning traffic, and small-model pricing determines whether that traffic is affordable.

## What are the market implications for Anthropic, OpenAI, and Meta?

Ember's launch puts Meta directly in the segment that Anthropic and OpenAI have used to capture agent workload volume. Claude Haiku 5.5 and GPT-6 Luna are the models Meta chose as comparison points, and both are positioned as fast, low-cost options for production use. Meta's reported results claim a better AutomationBench-AA score than either at a lower per-task cost.

The competitive consequence depends on independent validation. Meta is reporting its own benchmark results, and the company names Artificial Analysis as the benchmark source for the cost-per-task comparison. Buyers who have seen optimistic launch benchmarks before will likely run their own evaluations on representative workloads before shifting traffic from Claude Haiku 5.5 or GPT-6 Luna to muse-ember-1.0.

## What should teams evaluate before adopting Muse Ember?

Teams evaluating Muse Ember should start with a workload audit. The model's strongest claims are tied to cost per task and to specific benchmark categories, so a production test should measure the same things: input tokens, output tokens, retries, and task completion rate. The 1,048,576-token context window should be tested with the exact document types a team uses, because long-context performance can vary by format.

The API pricing tier is another test variable. Meta's standard pricing applies to requests up to 200,000 input tokens, according to Meta AI Research, which means teams using the full 1,048,576-token context window should confirm how longer inputs are billed. Output limits, multimodal input, and tool-call reliability should be evaluated inside the same harness where the model will run in production.

## What comes next for the Muse family?

Meta has not disclosed a roadmap for the Muse family in the launch post. The immediate availability of muse-ember-1.0 suggests the company is using the Meta Model API and Muse Code as the distribution channel for a fast release cycle. The role Meta assigns to Ember, as the executor that pairs with Muse Wonder or Muse Spark, implies that future Muse releases will continue to sharpen the planner-executor split.

The near-term signal for the market is cost, not raw capability. If Ember holds up in independent testing, it gives teams a cheaper execution layer for the high-volume work agents do all day. If it does not, the 1-million-token context window and multimodal input still make it a notable entry in a small-model segment that has become the default starting point for production AI systems.

## Sources

1. [Source for the Muse Ember launch, its 1,048,576-token context window, $0.10/$0.50 pricing, AutomationBench-AA scores, and per-task cost comparisons against Claude Haiku 5.5 and GPT-6 Luna.](https://research.meta.ai/blog/muse-ember)
2. [Source for the product description of Muse Ember as the efficient execution model in the Muse family.](https://dev.meta.ai/docs/muse-ember)
3. [Source for the launch positioning that Muse Ember achieves performance comparable to Claude Haiku 5.5 at roughly one-third of the cost per task while scoring higher than GPT-6 Luna across all reasoning settings.](https://www.facebook.com/AIatMeta/posts/today-were-introducing-muse-ember-our-smallest-and-most-efficient-model-yetmuse-/1427997929499631/)

---
Source: https://aiintelreport.com/frontier-models/meta-launches-muse-ember
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt