Meta Superintelligence Labs on Monday released Muse Ember, its smallest and cheapest Muse model, built for the repetitive work that AI agents run all day: tool calls, small coding tasks, routing, data extraction and long-document parsing. It costs $0.10 per million input tokens and $0.50 per million output tokens and is available through the Meta Model API and Muse Code, according to The Decoder. Meta says it delivers more intelligence per dollar than Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna.
What Muse Ember is
Ember is the fourth model under the Muse brand, following the mid-weight Spark in April, its open-weight twin Glimmer in August and the Wonder frontier model in October, TechCrunch reports. It accepts text, images, audio and video and handles up to one million tokens in one prompt, so an agent can read screenshots, charts and scanned documents or work from voice memos and meeting recordings without a separate model, in Meta’s words.
Meta designed it to pair with a larger model. TechCrunch quotes the company’s framing: “Muse Wonder plans, Muse Ember handles the volume.” In a customer-support setup, a heavier model could set the rules for an issue and Ember would then apply them across tens of thousands of cases.
How the numbers compare
Meta leaned on Artificial Analysis’ composite index to make its cost argument. On that test Ember scores 36 at about $0.14 per task, while the larger Muse Spark 1.3 scores 37 at $2.39 per task, the Decoder says. Meta claims Ember leads in most of its own tests at every reasoning level against Haiku 5.5 and GPT-6 Luna.
It is not ahead everywhere. On FrontierCode 2.0-AA, a test of real coding tasks, Haiku scores 34.9 percent against Ember’s 27.2 percent. At $0.10 per million input tokens, TechCrunch says Ember is also cost-competitive with open-weight rivals such as DeepSeek V4.1 Flash and GLM 5.3 Flash. The cost and index figures come from Meta’s own release material using a third party’s methodology, so treat them as vendor-presented until the leaderboard lists the model. Our Claude Haiku 5.5 page shows the competitor’s specifications.
What it means for developers and for Meta
Small models matter because agent workloads multiply calls: a single task can trigger dozens of tool invocations, and token price decides whether it is affordable at scale. A model priced at $0.10 per million input tokens makes high-volume routing and extraction cheaper to run, though quality on harder coding work still favors some rivals.
For Meta, Ember rounds out a family that now spans a small, a mid-weight and a frontier tier, alongside consumer features such as Spaces in the Muse app. Unlike Spark’s open sibling, the coverage we reviewed does not say that Ember has open weights, so we treat it as API-only. Meta has also been expanding its compute, including a $3 billion Nvidia investment.
What to watch next
- Whether Artificial Analysis and other leaderboards confirm Meta’s cost-per-task claims.
- Whether Meta releases an open-weight Ember variant, as it did with Spark and Glimmer.
- How Anthropic, OpenAI and Google price their small models in response.
- Developer feedback on Ember’s coding and tool-calling reliability.
- Usage of the planner-plus-worker pattern with Muse Wonder.
Key takeaways
- Muse Ember is Meta’s lightest Muse model, launched November 9.
- It costs $0.10 per million input tokens and $0.50 per million output tokens.
- It handles text, image, audio and video with a 1M-token context window.
- Meta cites a score of 36 on Artificial Analysis’ index at about $0.14 per task.
- Haiku 5.5 still beats it on a real-world coding test, 34.9 percent to 27.2 percent.
Frequently asked questions
What is Muse Ember used for?
Meta built it for repetitive agent work such as tool calls, small coding subtasks, routing, extraction and long-document parsing, ideally paired with a larger planning model.
How much does Muse Ember cost?
It is priced at $0.10 per million input tokens and $0.50 per million output tokens through the Meta Model API and Muse Code.
Is Muse Ember open source?
The reports we reviewed do not say its weights are open. Meta released the earlier Spark model’s twin, Glimmer, as open source, but Ember is described only as available through Meta’s API and Muse Code.
