Meta Muse Ember matches Claude Haiku 5.5 performance at one-third the cost per task

Meta Launches Muse Ember for High-Volume Agent Workloads

Meta Superintelligence Lab released Muse Ember on Monday, November 9, positioning it as the company's smallest and most cost-efficient AI model to date1

1

. The lightweight AI model targets repetitive agent execution tasks that AI systems perform throughout the day, including tool calls, coding subtasks, request routing, information extraction, and document parsing2

2

. CEO Mark Zuckerberg announced the launch on X, describing Ember as "our smallest model that delivers higher intelligence per dollar than others in its class"4

4

.

Muse Ember is the fourth model released under Meta's Muse brand, following the mid-weight Spark released in April, its open-source twin Glimmer in August, and the frontier model Muse Wonder in October1

1

. The proprietary AI model represents Meta's strategy to make advanced capabilities affordable enough to run across everyday products and services3

3

.

Cost Efficiency Breakthrough Against Claude Haiku and GPT-6 Luna

At $0.10 per million input tokens and $0.50 per million output tokens, Muse Ember matches the list prices of its closest competitors, Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna2

2

. However, the cost-efficient AI model delivers significantly better value according to independent benchmarking. On Artificial Analysis's Intelligence Index, Ember scored 36 at maximum reasoning effort, matching Claude Haiku 5.5's score while costing approximately $0.14 per task compared to Haiku's $0.414

4

.

Meta's Chief AI Officer Alexandr Wang emphasized this advantage on X, stating that "our teeny model muse ember is now on the efficient frontier, matching haiku at 3x cheaper cost per task, and smarter than luna"2

2

. GPT-6 Luna scored 27.8 at about 15 cents per task at its highest setting, though it could run for as little as 7 cents at lower settings with reduced performance4

4

.

The cost efficiency extends beyond standard competitors to open-source alternatives. Ember's input price undercuts Chinese rivals including DeepSeek V4.1 Flash, Z.ai's GLM-5.3 Flash, and MiniMax M3.1 Flash, all of which score lower on the Artificial Analysis index2

2

.

Designed for Agentic Workflows with Muse Wonder

Meta designed Muse Ember to pair with larger models like Muse Wonder in agentic workflows, creating a two-tier system where the frontier model handles planning while Ember executes high-volume tasks1

1

. In support ticket resolution scenarios, a heavier model might establish rules for handling particular issues, then deploy the lightweight AI model to apply those rules across tens of thousands of cases. As Meta summarizes: "Muse Wonder plans, Muse Ember handles the volume"1

1

.

This division of labor addresses a key challenge in AI agent deployment. A single support ticket can send an agent through multiple lookups, document reads, and tool calls before producing an answer, making cost management critical at scale4

4

. Ember is the first model Meta trained using Muse Wonder Full, demonstrating the integration between the two systems2

2

.

Multimodal Inputs and Extended Context Window

The AI model supports multimodal inputs including text, images, audio, and video, allowing agents to read screenshots, charts, scanned documents, voice memos, meeting recordings, and video content without requiring separate models1

1

. Meta lists a context window of 1,048,576 tokens with a maximum output of 524,288 tokens, providing enough capacity to process entire codebases or large document collections in a single call3

3

.

In early testing shared by Meta, cloud storage company Box used Ember to check contracts clause by clause against negotiation rules and assign overall risk ratings5

5

. The extended context window and document parsing capabilities make such workflow automation practical at scale.

Performance on Business Workflows and Coding Tasks

On AutomationBench-AA, which measures end-to-end business workflows, Muse Ember scored 37.8% compared to 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.53

3

. The test evaluates whether agents can fully complete tasks from start to finish across simulated business applications, retrieving records, making sequential tool calls, and updating information4

4

.

On Agents' Last Exam, Ember achieved 61.6%, narrowly exceeding Luna's 57.1% and Haiku's 57%3

3

. For coding subtasks, Ember scored 40.9% on Terminal-Bench 4.0 compared to 39.2% for Haiku and 16.4% for Luna3

3

. However, Claude Haiku performed better on FrontierCode 2.0-AA, which grades whether code changes are ready to merge, scoring 34.9% against Ember's 27.2%2

2

.

On OSWorld 2.1, which evaluates computer-use capabilities across desktop applications, Ember scored 69.5% using partial-credit scoring but fell to 33.8% when tasks were graded on an all-or-nothing basis4

4

. These mixed results suggest Ember performs best when given bounded jobs with clear checks rather than entire engineering tickets or unattended desktop workflows.

Pricing Structure and Model API Availability

The Meta Model API offers Muse Ember as muse-ember-1.0 through its Responses, Chat Completions, and Messages endpoints4

4

. At up to 200,000 input tokens, Meta charges $0.10 per million uncached input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Above 200,000 tokens, rates rise to $0.30, $0.012, and $1.50 respectively, with the higher rates applying to the entire request including output4

4

.

Artificial Analysis attributes part of Ember's benchmark cost advantage to cheaper cache reads and reports that Ember used roughly one-third as many output tokens as Haiku at maximum effort4

4

. This distinction matters when estimating production bills, as the same list price can yield very different costs once an agent starts reasoning, retrieving records, and carrying results from one step to the next.

Ember is also available through Muse Code, Meta's coding tool, and Oracle Cloud, with Google Cloud preview documents indicating text and image input support4

4

. Meta has kept Ember's model weights and parameter count private, maintaining its proprietary status like previous Muse releases2

2

.

Safety Framework and Future Development

Meta tested Ember against its safety framework covering chemical and biological weapons, cybersecurity, and loss-of-control scenarios. The company reported that Muse Ember remained below the framework's high-risk thresholds in every category2

2

. This evaluation process aligns with Meta's broader approach to AI safety as it scales deployment across consumer and enterprise applications.

The release supports Meta's Muse agent, which passed 5 million downloads in 22 days following its September launch5

5

. Recent updates include voice conversations and animated avatars rolling out in the United States. Meta indicated it already has even more cost-efficient models in training, suggesting continued focus on making AI agent deployment economically viable at scale5

5

. Watch for how developers integrate Ember into production workflows and whether the cost advantages hold as usage patterns evolve beyond benchmark scenarios.