Muse Ember Subagents: Fix the Max-Effort Default

Meta shipped Muse Ember on November 9, a small model aimed at the cheap end of an agent: tool-call subtasks, routing, extraction, small code fixes. It takes text, image, audio, video and PDF, has a 1M-token context window, and costs $0.10 per million input tokens and $0.50 per million output. That price is why people are wiring it in as the subagent behind a bigger planner model.

There is a catch that most of the launch-day posts skip. Per Meta's model page, the default reasoning effort for Ember is max. If you drop it into an existing subagent slot and change only the model string, every classification call will think as hard as the model allows, and you pay for those reasoning tokens as output. The second catch is the pricing tier: once a request goes over 200K input tokens, the whole request is billed at $0.30 / $1.50, not just the overflow.