Meta's Muse Ember costs $0.10 per million input tokens, and Simon Willison puts Claude Fable 5.2 at "80x that price"

What changed: On 9 November, Meta released Muse Ember at $0.10 per million input tokens and $0.50 per million output. It has a 1-million-token context. Meta built it for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction". Simon Willison notes that Claude Fable 5.2, released the same week at $8 and $40, is "80x that price".

Why it matters: I think most of what an agent costs comes from small steps it repeats, not from the hard reasoning. Running those steps on a frontier model is now the biggest cost you can cut. The agent's work stays the same.

What to do: Pull one week of token usage for your busiest AI workflow and split it by step. If classification, extraction and routing make up more than half the spend, move those steps to a small model. Compare 50 outputs side by side before you switch for good.