How to Use Muse Wonder
Last verified: October 28, 2026 · Sources: Meta announcements, Artificial Analysis
Quick answer
Muse Wonder is Meta's first closed-source, paid frontier model (released October 13, 2026). You can use it three ways: the Muse app (free download; $20/$100/month for higher limits), Muse Code ($50/$200/month, the only way to get Wonder Full), or the Meta Model API (model ID muse-wonder-1.0, $4/$20 per 1M tokens). It is best at long-lived agents and coding; weakest at factual knowledge.
What Muse Wonder is
A frontier LLM from Meta — and a strategic break from its open-source tradition: both Wonder and Wonder Full are closed and paid. Key specs:
| Spec | Value |
|---|---|
| Context window | ~1M tokens (1,048,576) |
| Input | Text, images, video, audio, PDF — in one context |
| Output | Text only |
| Reasoning effort | Adjustable up to max |
| Intelligence Index (AA v4.3) | Wonder 52 / Wonder Full 56 |
Step 1 — pick your access path
| If you want to… | Use | Cost |
|---|---|---|
| Chat and run everyday agents | Muse app (free download) | $0 to start; $20 or $100/mo for higher limits |
| Build coding agents on Wonder Full | Muse Code | $50 or $200/month |
| Call it from your own code | Meta Model API | $4/$20 per 1M tokens (cached $0.20) |
Details and the Full-vs-Max distinction: see subscription plans and API pricing.
Step 2 — check availability
- You must be 18+
- Not offered in China (including Hong Kong and Macau) or Russia; circumventing geo-controls is prohibited. If you are in those regions, look at alternatives like Gemini 4 Argon instead.
Step 3 — play to its strengths
What Wonder is genuinely built for, and what early users report:
Use it for
- Agents that should remember you — Active Memory carries context, preferences and decisions across sessions, not just within one chat
- Long-running tasks — Active Focus retries alternate paths when blocked instead of erroring out mid-task
- Coding agents — Wonder Full + Muse Code scores 68 on the coding-agent index (tied with Sonnet 5.5 + Claude Code)
- Multimodal documents — drop video, audio, PDFs and text into one ~1M-token context
Don't use it for
- Fact-heavy research — AA-Omniscience 30 vs Opus 5.5's 46; verify anything factual. For that workload see the Opus 5.5 comparison
- Short-range precision tasks — a noted weak spot in independent testing
- Generating images or video — output is text only
Wonder vs Wonder Full: which should you use?
Start with the base Wonder (free app tier or cheap API calls). Upgrade to Full via Muse Code only if you run coding agents or hit capability walls: the Intelligence Index gap is 52 → 56, and the coding-agent score (68) requires the Muse Code pairing. Remember: setting the base model's reasoning effort to "max" does not equal Full.
How it stacks up
- vs Claude Opus 5.5 — the capability leader at the same token price
- vs Gemini 4 Argon — half the price, same context size
- vs GPT-6 — 2.5x more expensive than Wonder
- vs Spark 1.3 — if you're upgrading from Meta's previous generation
Data verified October 28, 2026. Unofficial guide; see about.