November 9, 2026 — Meta Superintelligence Labs released Muse Ember on Monday, its lightest and most cost-efficient AI model yet, priced at $0.10 per million input tokens and built for the repetitive work AI agents do all day. Why it matters: Ember is a bet that the real cost of agents is not the one smart decision, but the thousands of small ones — and Meta wants to own that volume.
What happened
Ember is the fourth model in Meta’s Muse family, after the mid-weight Muse Spark in April, its open-source twin Glimmer in August, and the frontier Muse Wonder in October. Meta describes Ember as designed for complex reasoning execution and the tasks agents repeat constantly: tool calls, coding subtasks, routing, extraction, and long document parsing. It is available through the Meta Model API (as muse-ember-1.0) and inside Muse Code, with closed weights.
The details
Pricing is the headline: $0.10 per million input tokens and $0.50 per million output tokens — a rate TechCrunch notes is cost-competitive even with open-source models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta is pitching cost-per-task rather than raw capability: on the Artificial Analysis Intelligence Index, Meta puts Ember at a score of about 36 at roughly $0.14 per task, comparable in score to Anthropic’s Haiku at nearly three times the cost per task. The model takes text, images, audio, and video inputs — Meta says agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model — and it offers a 1-million-token context window. Meta also reports Ember as its first model trained using Muse Wonder Full, and says it can operate desktop applications from screenshots. These benchmark figures are Meta’s own; on the FrontierCode 2.0 coding test Meta reports, Ember trails Haiku, so the launch numbers should be read as a vendor’s opening claim, not a settled ranking.
Built to work under a bigger model
Ember is designed to pair with a larger model in an agentic harness: the heavy model plans, the light model executes at scale. Meta’s own framing is that Muse Wonder plans and Muse Ember handles the volume. TechCrunch’s example is a customer-support agent where the larger model sets the rules for handling an issue type, and Ember then applies those rules across tens of thousands of cases.
Why it matters
Agent economics are quietly becoming the real AI race. A frontier model answering one hard question is cheap; an agent making ten thousand tool calls a day is not. By pricing Ember at open-source levels while keeping it inside its own API and coding tools, Meta is trying to make the workhorse layer of the agent stack a Meta product. For builders, the practical takeaway is the pattern, not just the model: plan with the smartest model you can afford, execute with the cheapest one that holds accuracy — everything AI, tested, means checking whether Ember’s cost-per-task claims survive your own workload before you route production traffic to it.
Frequently asked questions
What is Muse Ember?
Muse Ember is Meta Superintelligence Labs’ lightest and cheapest AI model, released November 9, 2026, and built for high-volume agent execution tasks such as tool calls, routing, extraction, and document parsing.
How much does Muse Ember cost?
Meta prices Ember at $0.10 per million input tokens and $0.50 per million output tokens through the Meta Model API — about $0.14 per task on the Artificial Analysis composite benchmark Meta cites.
Where can developers use Muse Ember?
Ember is available through the Meta Model API under the model name muse-ember-1.0 and inside Muse Code. Its weights are closed.
How is Ember different from Muse Wonder?
Wonder is Meta’s frontier model, meant for planning and hard reasoning. Ember is the smaller execution layer designed to carry out high volumes of routine agent steps cheaply, often under a larger model’s direction.
Sources: TechCrunch, Meta Superintelligence Labs announcement coverage (November 9, 2026).