Meta Muse Wonder and Wonder Full: Price, Access, and Benchmark Reality Check
Key Takeaways
- Meta has released Muse Wonder and Muse Wonder Full, its first genuinely frontier-class AI models, launched October 12, 2026.
- Wonder is available on the Meta Model API and in Muse Code at $4/$20 per million input/output tokens. Wonder Full is subscribers-only and not on the public API.
- Both models offer 1M-token context and cross-chat memory for agentic workloads. Both have closed weights.
- Independent benchmarks score Wonder Full at 56 on the Artificial Analysis index, versus 58 for Claude Opus 5.5. Competitive, but not the outright lead.
- This is a strategic shift: Meta is now competing directly with the top closed labs instead of only releasing open models.
Meta has released Muse Wonder and Muse Wonder Full, its first genuinely frontier-class AI models, launched October 12, 2026. Positioned for agentic workloads with cross-chat memory, they mark Meta’s step from capable open models into direct competition with the top closed labs.
Here is what they cost, where you can use them, and what independent benchmarks actually show versus Meta’s claims.
What Are Muse Wonder and Wonder Full?
Wonder is the standard frontier model. Wonder Full is the larger, more capable variant. Both are designed around agentic use, which means models that can work across extended sessions with persistent memory of prior conversations.
| Muse Wonder | Muse Wonder Full | |
|---|---|---|
| Availability | Meta Model API + Muse Code | Subscribers only (not on the public API) |
| API pricing | $4 / $20 per 1M input/output tokens | Not API-available |
| Context | 1M tokens | 1M tokens |
| Weights | Closed | Closed |
Note the pricing: $4/$20 per million tokens puts Wonder exactly in line with top-tier frontier models like Claude Opus 5.5. Meta is pricing for the top of the market, not undercutting it.
What “Cross-Chat Memory” Means in Practice
Most chatbots start each conversation with a blank slate. You tell it your preferences on Monday, and by Wednesday it has forgotten them. Cross-chat memory changes that. The model carries context from earlier conversations into new ones, so it can build on what you discussed before without you repeating yourself.
For agentic workloads, this matters a lot. An agent working on a week-long project should remember the decisions you made on day one. It should know your coding style, your constraints, and what you already tried. Without persistent memory, every session starts over. With it, the model behaves more like a colleague who was in the room last time.
There is a trade-off worth noting. Memory that persists across chats is convenient, but it also means the model holds more information about you over time. Check what memory controls Meta provides, and review what is stored if you work with sensitive material. Convenience and privacy pull in opposite directions here, as they do with every assistant that remembers you.
Benchmarks: Meta’s Claims vs Independent Scores
- Meta claims these are its “most capable yet” models with leading benchmark results. Treat vendor claims as marketing until independently confirmed.
- Independent data: The Artificial Analysis index scores Wonder Full at 56, versus 58 for Claude Opus 5.5. That is genuinely competitive, within striking distance of the best, but not the outright lead Meta’s framing suggests.
A two-point gap on a composite index is small but real. It means Wonder Full handles most frontier tasks about as well as the leaders, while trailing slightly on the hardest evaluations. For everyday building, that difference may never show up. For teams chasing the absolute best scores on difficult reasoning, it is the reason to test both before committing.
How This Fits Meta’s AI Lineup
Meta’s naming is getting crowded, so here is the map:
- Muse (AI agent), the consumer assistant we covered in our Muse AI guide
- Muse Image, the image generation model (guide here)
- Muse Spark Open, the open-weight model for local running
- Muse Wonder / Wonder Full, the new closed frontier models (this article)
Wonder is the engine that will likely power future versions of the Muse agent, so even if you never touch the API, this release shapes the assistant millions use.
How to Access Muse Wonder
There are two practical routes. Developers can call Wonder through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Muse Code users get Wonder directly inside the coding tool. Wonder Full, the larger variant, is available to subscribers but not on the public API.
If you want to understand the agentic workloads these models target, our AI agents explainer is a useful primer.
Wonder vs the Competition
The honest picture: Wonder enters a market with established leaders. Claude Opus 5.5 scores 58 on the Artificial Analysis index against Wonder Full’s 56. Pricing is matched at $4/$20, so Meta is not buying market share with discounts. It is asking developers to choose Wonder on capability and ecosystem, not price.
That is a bold move for a first frontier release. It signals that Meta believes the model can stand next to the best without a price cut. For developers, the practical response is simple: test it on your own tasks.
The Strategic Shift: Why “Closed” Matters
Meta spent years as the champion of open AI. Its Llama models were free to download and run. Wonder breaks that pattern. Both Wonder models have closed weights. You can call them through Meta’s API, but you cannot download or self-host them.
This is a real strategic turn. Open weights build goodwill and a developer community. Closed frontier models capture API revenue and keep the best capabilities in-house. Meta is now doing both: Muse Spark Open serves the open side, while Wonder serves the closed frontier.
Pros and Cons
Strengths: genuinely competitive frontier performance; 1M-token context; cross-chat memory for long projects; available in Muse Code without API setup; likely to improve the consumer Muse agent over time.
Weaknesses: closed weights, a break from Meta’s open tradition; top-tier pricing with no discount; Wonder Full not on the public API; independent scores trail the very best by a small margin; vendor benchmark claims unverified.
Who Should Care?
- Developers comparing frontier APIs. Wonder is now a legitimate third option alongside the usual leaders, at matching prices.
- Muse Code users. Wonder is available there directly, so trying it costs nothing extra.
- Anyone tracking the AI race. Meta going closed-frontier is a strategic shift worth watching.
- Teams already in Meta’s ecosystem. If you use Meta’s tools, Wonder deepens what you can build without switching providers.
Frequently Asked Questions
When was Muse Wonder released?
October 12, 2026.
How much does the Muse Wonder API cost?
$4 per million input tokens and $20 per million output tokens. Wonder Full is not on the public API.
Is Muse Wonder open source?
No. Both Wonder and Wonder Full have closed weights. (Muse Spark Open is the separate open-weight release.)
What is the difference between Wonder and Wonder Full?
Wonder Full is the larger, more capable variant. It is available to subscribers but not through the public API. Developers building on the API use Wonder.
Can I run Muse Wonder locally?
No. The weights are closed, so there is no download or self-hosting option. For local running, look at Muse Spark Open instead.
Will Wonder improve the Muse assistant I already use?
Likely, over time. Wonder is the engine expected to power future versions of the Muse agent, so improvements should flow into the consumer assistant.
