Home / News

Meta's Muse Ember makes the volume layer of AI agents almost free to run

Model fees are becoming the smallest line in the budget for an AI assistant in a small business: Meta's new Muse Ember costs $0.10 per million input tokens. Most of the money in a small-business AI project now goes into building, testing and maintaining the workflow, while the model cost of routine work like extraction and routing keeps falling.

According to TechCrunch, Meta Superintelligence Lab built Ember for the repetitive work agents do all day: tool calls, routing, data extraction and long document parsing. It also reads screenshots, scans and audio without a separate model. Meta pitches it as a pair for its heavy Muse Wonder: "Muse Wonder plans, Muse Ember handles the volume."

What changes for small-business automation

The split itself is sound, and I already build agents this way: a strong model like Claude writes the rules, and a cheap one applies them to hundreds of invoices, support tickets or scanned receipts. Ember makes that volume layer cheaper. Three caveats, though. The "smarter than Haiku" claim comes from Meta's own chart. The article only gives the input price. And the model is closed, so your data goes to Meta's API, which rules it out for clients who must keep documents in Russia or on their own servers. Even before Ember, the model bill was rarely what drives a project's cost - integrations and edge cases are.

Why it matters: before switching to a cheaper model, run it on 50-100 of your real documents and compare its errors with your current setup. A model that misreads invoices costs more in manual corrections than it saves in tokens.

Source: TechCrunch