Summary: Cursor’s docs now list Muse Wonder, Meta’s flagship model and the successor to Muse Spark 1.3, positioned for agentic coding work — long tool chains, multi-step debugging, and large repositories. For builders, the interesting knobs are its six effort levels and a 300k-token context window that stretches to 1M in Max Mode without a long-context surcharge. One catch: it bills as a third-party model, not from your plan’s included usage.
What Muse Wonder is built for
According to Cursor’s model documentation, Muse Wonder is Meta’s flagship model aimed squarely at agent workloads: long tool-use chains, multi-step debugging, and large-repository tasks. That’s a different pitch from a general chat model — the target is the loop where an agent plans, calls tools, reads results, and keeps going without losing the thread.
In Cursor it gets access to all agent tools, so it can use the same toolset as the models you’re already routing to. If you’ve followed the Muse family on this blog, this continues a line that started with smaller local options like Meta’s Muse Glimmer, a 30B open model for local agents — Wonder is the bigger, hosted end of that family.
The effort dial is the real decision
The standout feature for day-to-day use is the six-level effort control: minimal, low, medium, high (the default), extra high, and max. Higher settings spend more reasoning tokens per step and score better on CursorBench; the docs explicitly suggest dropping to low or minimal for simpler tasks to get faster, cheaper responses.
This matters because effort levels are where agentic coding bills quietly balloon. A model that reasons hard on every step is great on a gnarly multi-file refactor and wasteful on a rename-and-run-tests loop. If you’re routing between models anyway, Muse Wonder’s dial gives you a second lever: keep the model, lower the effort. I’ve written before about treating model settings as spending controls in how to spend Gemini 4 Argon wisely in Cursor, and the same discipline applies here.
Context, images, and billing
Muse Wonder ships with a 300k-token context window, expanding to 1M tokens in Max Mode at the same per-token rates — no long-context multiplier. It also accepts image input, including screenshots returned by tools, which is increasingly table stakes for agents that need to read UI or rendered output.
On cost: this is a third-party model, so it isn’t part of the usage included with Cursor plans. On The X Subscription it draws from prepaid credit; on legacy plans it pulls from the Other Models pool. The exact per-token rates appear in a pricing table that the supplied RSS summary does not specify. One confirmed detail: cached input bills at $0.20 per million tokens, with no cache-write charge. That’s worth knowing if your agent workflow re-sends large stable context — caching is where the savings live.
One admin step before you can try it
If you’re on a Teams or Enterprise plan and configured model access before Meta became available, Muse Wonder won’t just appear. An admin needs to enable the Meta provider first; only then does the model show up for team members. It’s a small step, but it’s the kind of thing that makes a model seem “missing” when it’s really just gated behind provider settings.
On privacy, Muse Wonder runs under Cursor’s standard provider terms, with Meta’s zero data retention policy applying on Meta’s side. Worth a skim before pointing it at proprietary code.
Practical takeaway
If your agents already run in Cursor, Muse Wonder is a low-friction addition: enable the provider, try it at default high effort on your hardest debugging task, then dial down to low or minimal on routine work and watch the token spend. The effort control is the feature to build a habit around — the model choice matters less than matching reasoning spend to task difficulty.
Sources
View the original post on X