Meta Launches Muse Ember to Cut AI Agent Costs to 14 Cents Per Task

Quick Facts

• Muse Ember is priced at $0.10 per million input tokens and $0.50 per million output tokens, with cached inputs dropping to $0.004 per million tokens.
• The model scores 37.8% on AutomationBench-AA business workflows, nearly tripling Anthropic's Claude Haiku 5.5 at 13.9% and OpenAI's GPT-6 Luna at 20.2%.
• Ember is the first model distilled from Muse Wonder Full and supports text, images, audio, and video within a 1 million token context window.

Meta on Tuesday introduced Muse Ember, its smallest and most cost-efficient AI model, built to handle the repetitive execution work that AI agents perform at scale. The model is available now through Muse Code and the Meta Model API, and positions itself directly against Claude Haiku 5.5 and GPT-6 Luna on cost per completed task.

Muse Ember costs about $0.14 per Artificial Analysis task. By comparison, Muse Spark 1.3 runs $2.39 per task. Haiku 5.5 costs roughly $0.41 per task.

Cached input pricing falls to $0.004 per million tokens for contexts up to 200,000 tokens. Meta says that figure is 60% cheaper than its nearest competitors on cached input, according to Artificial Analysis data.

Where It Fits in the Muse Family

Ember is the fourth model in Meta's Muse lineup, following Muse Spark, Muse Glimmer, and Muse Wonder. Mark Zuckerberg announced Muse Wonder and Muse Wonder Full earlier, calling them Meta's entry to the frontier model tier. Zuckerberg said on X that Muse Spark was the first family of models from Meta Superintelligence Labs and that Muse Wonder represents a larger and more advanced class.

Ember is the first model trained using Muse Wonder Full. Meta describes the division of labor as: Muse Wonder plans, Muse Ember handles the volume.

In practice, this means a frontier model like Muse Wonder breaks a user request into smaller tasks and routes them to Ember agents for execution. Meta's developer blog gives a concrete example: Muse Wonder planned a port of 16 lexers from the Python library Pygments to Rust, and Ember wrote the code.

Benchmark Performance

On the Artificial Analysis Intelligence Index, Ember scores 36 at maximum effort. That matches Haiku 5.5 and beats GPT-6 Luna's 28.

Benchmark results across agent-focused tasks show Ember leading on most execution categories. On Terminal-Bench 4.0, Ember scores 40.9% versus Haiku's 39.2% and Luna's 16.4%. On OSWorld 2.1, which tests computer use, Ember scores 69.5% against Haiku's 68.9% and Luna's 50.5%.

The Coding Agent Index shows Ember at 46%, Haiku at 37%, and Luna at 31%. The one clear loss: FrontierCode 2.0, which measures whether real engineering changes are mergeable. Haiku wins that category with 34.9% to Ember's 27.2%.

Ember also scores 21.1% on AA-Omniscience, declining 64% of knowledge questions to avoid hallucinations. That tradeoff limits its use in question-answering applications but reduces the risk of confident wrong answers in automated pipelines.

Technical Specs

Muse Ember supports text, images, audio, and video in a single model, removing the need for separate vision or audio models in a stack. The context window holds up to 1 million tokens, with output up to 512,000 tokens per request.

For inputs above 200,000 tokens, pricing rises to $0.30 per million, with output above that threshold at $1.50 per million. The model is available as muse-ember-1.0 through the Meta Model API.

Business Case for Founders

The economics matter for any team building agent-based products. At $0.14 per completed Artificial Analysis task, Ember makes it feasible to run agents on routine operations — support ticket routing, document extraction, code generation subtasks — without the cost of a frontier model on every call.

Meta's framing targets the same workflow that most B2B SaaS teams are already building: an orchestrator model that plans, and cheaper worker models that execute. Ember is designed to occupy that worker role across high-volume, multimodal pipelines at a price point Meta says is low enough to embed in every product, for every user.