One signal

Update What changed

Meta says Muse Ember is available through both Muse Code and the Meta Model API, with support for the Responses API. Its listed prices are $0.10 per million input tokens, $0.50 per million output tokens and $0.004 per million cached input tokens.

The model has a one-million-token context window and can produce up to 512,000 tokens of output. Meta says it handles text, images, audio and video, so agents can process material such as screenshots, charts, scanned documents and recordings without a separate model for each format.

Meta positions Ember as a smaller model for repeated agent tasks, including tool calls, coding subtasks, routing and document parsing, alongside its larger Muse Wonder model.

Meta also presents benchmark comparisons suggesting Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna.

Sources and evidence

• Introducing Our Smallest, Most Efficient Model: Muse Ember - Meta AI: Meta says Muse Ember is available through Muse Code and the Meta Model API, costs $0.10 per million input tokens, $0.50 per million output tokens and $0.004 per million cached input tokens, and supports multimodal inputs with a one-million-token context window and up to 512,000 output tokens.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Meta AI chief Alexandr Wang describes Muse Ember as Meta’s smallest and most efficient model, built for work that AI agents repeat throughout the day.

Wang claims Ember delivers more intelligence per dollar than Haiku, Luna and other “flash” models across reasoning settings. That is a broad comparison, and the claim comes from Meta’s executive rather than an independent benchmark.

He also says Ember costs a third as much as those alternatives, or less. For developers weighing repeated agent calls, that could make the model’s economics more notable than its size. Which matters more when choosing an agent model: a lower stated price, or independently comparable performance evidence?

Sources and evidence

• alexandrwang on X: 1/ announcing our smallest, most efficient little model is here: muse ember 🔥 built for work that agents repeat all day, it gets you more intelligence per dollar: Alexandr Wang says Muse Ember is Meta’s smallest and most efficient model, is designed for repeated agent work, and delivers more intelligence per dollar than Haiku, Luna and other flash models across reasoning settings at a third of their cost or less.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Meta says Muse Ember is the first model trained using Muse Wonder Full, the larger model it introduced as part of its Muse line. That adds a detail about how the new lightweight model was developed, beyond its role handling repeated agent tasks.

Meta also says it assessed Ember for chemical and biological, cybersecurity and loss-of-control risks under its safety framework, and that its evaluations stayed below the framework’s high-risk thresholds. Those are Meta’s reported results, not an independent safety finding.

For agent use, Meta says it tested whether Ember follows user intent, handles sensitive information appropriately, resists hidden instructions in web pages, documents or tool outputs, and seeks human input before high-impact actions. The company says the model received the same core safety training and testing as its other models.

Sources and evidence

• Meta releases Muse Ember, a new lightweight model for agent execution: Meta says Muse Ember was trained using Muse Wonder Full and that its safety evaluations remained below its framework’s high-risk thresholds; it also describes checks of agent behaviour, including resistance to hidden instructions and requests for human input before high-impact actions.

Independent WittyWires Watcher; not an official account or feed.