Meta introduced Muse Ember on November 9 for repeated agent tasks; it is available through Meta Model API and Muse Code.
The big change
- The change: A smaller model joins Meta’s lineup.
- Why it matters: Small charges accumulate across many agent steps.
- What to watch: Token prices alone do not establish cost per successful task.
Understanding the bill
Developers must decide where to delegate and check results before acting.
According to Meta’s pricing documentation, prompts up to 200,000 input tokens cost $0.10 per million input, $0.004 cached input and $0.50 output. Above 200,000, the respective rates are $0.30, $0.012 and $1.50. Cached tokens count toward the threshold; the higher tier applies to the entire request.
At the lower tier, one million uncached input tokens across qualifying requests plus 100,000 output tokens cost $0.15 in model tokens. Retries and tool use can increase the final bill.
What an independent test measures
Vals AI reported 60.48% for Ember on its Vals Index at maximum reasoning and about $0.15 per test in its November 9 release note. Its methodology combines finance, coding, legal and tax benchmarks weighted by US economic output. The finance component includes multi-step financial reasoning and spreadsheet modeling; other components cover different kinds of work. The aggregate describes selected tests, not the accuracy of a particular business workflow.
The Vals test cost and token bundle measure different things. Teams can test representative inputs, failure rates and complete costs before deciding which actions to delegate. BIG CHANGE has not tested Ember.
