Meta releases Muse Ember at 10 cents per million input tokens

Meta released Muse Ember on 9 November, the smallest and cheapest model in its Muse family. The model is live on Muse Code and the Meta Model API with the name muse-ember-1.0. Meta pitches it for tool calls, coding subtasks, routing, extraction and long document parsing.

The price and the context window

Muse Ember costs 10 cents per million input tokens and 50 cents per million output tokens. Cached input is 0.4 cents per million tokens. The context window is 1 million tokens. The model can return up to 512,000 tokens in one response. It supports the Responses API, which is Meta's streaming tool-use interface.

What Meta compares it to

Meta names two rivals in the post. One is Anthropic's Claude Haiku 5.5. The other is OpenAI's GPT-6 Luna. On the Artificial Analysis Intelligence Index, Meta says Muse Ember scores 36. The company says that matches Haiku's best score at about a third of the cost. On the same index, Meta says Muse Ember beats every Luna setting. The numbers are Meta's own, measured on Meta's own test runs.

The scores Meta reports

Meta lists ten benchmarks for Muse Ember. AutomationBench-AA: 37.8 percent, against 13.9 percent for Haiku and 20.2 percent for Luna. OSWorld 2.1, a desktop automation test: 69.5 percent, against 68.9 percent for Haiku and 50.5 percent for Luna. Terminal-Bench 4.0: 40.9 percent. Agents' Last Exam: 61.6 percent. FrontierCode 2.0-AA: 27.2 percent. Coding Agent Index-AA: 46. PerceptionBench: 80.6 percent. GDP.pdf-AA: 22.8 percent. Personal Superintelligence Index: 46.8 percent. Each figure is Meta's own.

The named use cases

Meta names three partner tests in the post. Box says Muse Ember can score a contract against a company's negotiation playbook, clause by clause. The post also lists candidate screening and code file reading. Testing and targeted fixes are on the list too, along with desktop automation. These workloads run the same prompt many times, and the price per call is most of the bill.

What a reader can do today

A team running Haiku 5.5 on routing, extraction or document reading can test Muse Ember on the same prompts. The 1 million token context lets it read long contracts and whole call transcripts in one request. The same context fits most codebases too. The score for intelligence (36) is Meta's own. A third-party retest may give a different number. Teams with their own evals should run them before moving a production path.

Source

• Meta Research, Our Smallest, Most Efficient Model: Muse Ember
• The New Stack, Meta launches Muse Ember to take on Claude Haiku and GPT-6 Luna