Muse Wonder 1.0: Build a Cited Research Agent in Python

Meta shipped Muse Wonder on October 12, and for the first time its top model is something you can call from the same OpenAI-style client you already use. The API model is muse-wonder-1.0, with a 1M-token context window, $4 per million input tokens, $20 per million output tokens and built-in web search at $2.50 per 1,000 queries. The bigger sibling, Muse Wonder Full, is not in the API at all. It only exists inside Muse Code subscriptions.

That pricing split is the interesting part for builders. Wonder costs roughly three times what Muse Spark 1.3 does ($1.25 in, $4.25 out), so you do not want to point every request at it. A research agent is a good fit for the expensive model: few calls, lots of reasoning, and the output is only useful if you can trust it. So this guide builds exactly that, a research agent that answers with numbered citations, plus a second pass that re-reads every cited page and grades each claim.