Meta launches Muse Ember, a small model at $0.10 per million input tokens
Meta has introduced Muse Ember, which it describes as its smallest and most efficient model. The company says it is designed for complex reasoning execution and for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing.
The headline claim is about cost. Meta says Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting. It says the model matches Haiku's best score at about a third of the cost and beats every Luna setting. Meta also says Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard, meaning that no model scores higher for less per task. It puts the cost at about $0.14 per Artificial Analysis task. The post points to a separate report for the evaluation details; no scores are given in the post itself.
Muse Ember is available now on the Meta Model API and on Muse Code as muse-ember-1.0, and it supports the Responses API. Pricing is $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. The context window is 1M tokens, with up to 512K output tokens, which Meta says lets it take in a full codebase or document set in a single call. Muse Ember, Muse Spark and Muse Wonder are also now available on Google Cloud and Oracle Cloud.
Meta describes three areas of use. For agents, it says most tokens go to many small, structured steps, and that Ember can carry a business workflow end to end: pulling records, calling tools in sequence and updating systems across apps. For coding, it is aimed at subtasks an agent sends off hundreds of times a session, such as reading files, running tests and making targeted fixes; for the hardest problems Meta suggests pairing it with Muse Wonder, where Wonder plans and Ember handles the volume. On multimodality, Meta says Ember natively understands text, images, audio and video, so agents can read screenshots, charts and scanned documents, work from voice memos, meeting recordings and video without a separate model, operate a computer directly from screenshots and pick out fine visual detail.
On safety, Meta says it evaluated Ember under its Meta Superintelligence Scaling Framework, including chemical and biological, cybersecurity and loss-of-control risks, and found it below the framework's high-risk thresholds across the categories assessed. Meta adds that, though small, the model received the same core safety training and testing. It also tested how the model behaves when acting on a user's behalf: whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages, documents or tool outputs that try to redirect it, and seeks human input before high-impact actions.
Meta places Ember in a three-model lineup: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Ember for repeatable coding and agentic tasks at scale. It is the first model trained using Muse Wonder Full. Meta says it has even more efficient models in training. In early testing, it says, unnamed partners used Ember on repetitive agent tasks, from reviewing contracts to writing code.
Key facts
- Muse Ember is available as muse-ember-1.0 on the Meta Model API and Muse Code, at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens.
- It has a 1M-token context window and up to 512K output tokens, and supports the Responses API.
- Meta claims more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, matching Haiku's best score at about a third of the cost.
- Meta puts the cost at about $0.14 per Artificial Analysis task and says no model on the full leaderboard scores higher for less per task.
- It is the first model trained using Muse Wonder Full, understands text, images, audio and video natively, and is now also on Google Cloud and Oracle Cloud alongside Muse Spark and Muse Wonder.
Why it matters
Meta is pitching Muse Ember on price per unit of capability rather than on top-end performance. It is the third model in a lineup with Muse Wonder (hardest problems) and Muse Spark (coding and agentic workflows), and the first trained using Muse Wonder Full. Meta frames the goal as frontier capability "at a price that scales" and says more efficient models are already in training.
Who it affects
Developers building agents that make many small calls (tool use, routing, extraction, long document parsing, coding subtasks) are the target, since Meta says those steps eat most of an agent's tokens. Teams that need a multimodal model reading screenshots, audio or video, or one that can take a whole codebase in a 1M-token window, are also in scope. The competitors Meta names for comparison are Claude Haiku 5.5 and GPT-6 Luna.
How to use it
The model ID is muse-ember-1.0, available on the Meta Model API and Muse Code, with Responses API support. Meta lists $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, with a 1M-token context window and up to 512K output tokens. Ember, Spark and Wonder are also on Google Cloud and Oracle Cloud. For hard coding work, Meta suggests a split: Muse Wonder plans, Muse Ember handles the volume.
How solid is it
This is a first-party launch post from Meta, and every performance and cost claim is Meta's own. The comparisons with Claude Haiku 5.5 and GPT-6 Luna are stated in words only; the post gives no benchmark scores or Artificial Analysis index values and refers readers to a separate report. The competitors' prices are not stated, so the "about a third of the cost" figure cannot be checked from the post. The partners who tested the model are not named and no partner results are given.
Risks and caveats
Meta's safety conclusion, that Ember stayed below the high-risk thresholds of its own Meta Superintelligence Scaling Framework, comes from Meta's own evaluations. The post does not say which reasoning settings were compared or how the cost per Artificial Analysis task was computed beyond the quoted figures. It gives no parameter count or architecture details, and no statement on open or downloadable weights. Claims of efficiency-frontier status rest on a leaderboard snapshot that Meta describes but does not show.
“a cost low enough to run in every product, for everyone”
— Meta, Muse Ember announcement