Meta launched Muse Ember on November 9, 2026, giving businesses and application teams a new model for repeated AI tasks such as sorting support tickets and extracting document fields. It is available through Meta Model API and Muse Code, the company's developer service and coding tool.
Where Ember fits
Meta's model-selection guidance gives Ember a supporting role: a larger model can plan the work while Ember handles repeated, narrowly defined steps, including delegated coding tasks. For difficult coding problems to solve in one attempt, knowledge-heavy questions or visual coding, Meta recommends Muse Spark or Muse Wonder instead.
The model documentation lists text, image, audio, video and PDF inputs, with text output. Those capabilities make document-processing tasks candidates for evaluation; they do not establish how accurately Ember will handle a particular business's records.
When a long request costs more
Meta's pay-as-you-go rates depend on request size. Tokens are the units used to count model input and output. All prices below are US dollars per million tokens.
Billed usage | Up to 200,000 input tokens | Over 200,000 input tokens |
|---|---|---|
Uncached input | US$0.10 | US$0.30 |
Cached input | US$0.004 | US$0.012 |
Output | US$0.50 | US$1.50 |
The input total selects the tier for the entire request, including its output. Cached (reused) input and retained conversation history count toward that total, so a short new message can still fall into the higher tier.
Calculated example: assume one request with uncached text and 10,000 billed output tokens, including internal reasoning. At 200,000 input tokens, the charge is US$0.020 + US$0.005 = US$0.025. At 200,001, it is US$0.0600003 + US$0.015 = US$0.0750003, or about 7.5 cents.
This calculation excludes tools, retries, infrastructure and human review. It illustrates token billing, not a completed job's total cost.
There is also output you do not see. Meta says Ember defaults to its maximum reasoning setting, and internal reasoning tokens are billed as output. A short answer is therefore not necessarily a small output bill.
What to check before switching
For buyers, a sensible first evaluation is one task with a known correct answer: whether a ticket was categorised correctly, or whether an extracted field matches its document. Compare spending per accepted result, including failed attempts and review, rather than token prices alone.
Check data settings too. Ember's API access is Standard mode only; Meta says that mode does not use your content for training. This is not a promise of zero retention. Muse Code subscriptions default to Contributor mode, which permits training on submitted content and can be turned off. Subscription availability and benefits may vary by region.
For the broader budgeting question, RohitAI's coverage of tracking AI spending across teams and agent sessions explains the distinction between allocating costs and investigating where they build up.