Muse Wonder is Meta’s first frontier-class AI model, and it launched on October 12, 2026. Developers can use it through the Meta Model API today, while a bigger sibling is limited to Muse Code subscribers. Meta says it matches OpenAI and Anthropic, but those claims come from Meta’s own testing.
Quick facts before the details:
- Two models launched: Muse Wonder and Muse Wonder Full.
- Muse Wonder API pricing is $4 per million input tokens and $20 per million output tokens.
- Muse Wonder Full is only for Muse Code subscribers, with no API access.
- No independent testing had confirmed Meta’s benchmark claims at launch.
The details below come from TechCrunch and Axios, plus a more critical look from Fello AI.
What Muse Wonder is, and how it differs from Wonder Full
Meta’s Superintelligence Lab built both models. They sit above the lighter Muse Spark and aim at harder jobs, like coding and long-running agent work, according to TechCrunch.
Muse Wonder is the one most people can actually try. Muse Wonder Full is the stronger version, but Meta keeps it inside Muse Code. Both are meant to power Meta’s Muse assistant over time, starting with paid subscribers.
If you want background on that assistant, see our earlier post on the Meta Muse AI agent.
Muse Wonder price and who can use it
Access depends on the model. This table shows what the launch coverage says.
| Model | Where you can use it | Price or plan |
|---|---|---|
| Muse Wonder | Meta Model API and Muse Code | $4 input and $20 output per million tokens |
| Muse Wonder Full | Muse Code only, no API | Muse Code subscribers |
| Muse assistant | Coming later, paid users first | Not announced |
TechCrunch says the API price is roughly in line with Opus 5.5 and GPT-6 Sol. Fello AI adds that Wonder Full sits on Muse Code’s $50 and $200 monthly plans, though we couldn’t confirm that on Meta’s side.
How good is it? Meta’s numbers versus outside checks
Meta says Muse Wonder and Wonder Full score on par with Fable 5.1 and GPT-6 Astra. It also claims Muse Wonder beats rivals on MemoryBench. Those figures sit in Meta’s own model card.
Here’s the catch. TechCrunch notes that it found no independent evaluations at launch. Axios says the same, calling the real-world test an open question.
Fello AI offers a sharper view. It says Meta’s chart uses a method that favors its own column, and that two winning rows are Meta’s own benchmarks. It also cites Artificial Analysis scores that put Wonder Full at 56 and Muse Wonder at 52, below Opus 5.5 at 58. Treat that as a single outside reading, not a final verdict.
The practical point: the model you can call through the API is the weaker one, at a price that matches Opus 5.5. That’s worth testing before you switch anything.
Active Memory and Active Focus: what Meta is pitching
Meta highlights two features for agent-style work. Here’s what each is meant to do, in Meta’s words as relayed by Axios:
- Active Memory keeps information from your past chats and brings it back when it’s relevant.
- Active Focus keeps an agent working toward a goal and adjusting when it hits a wall.
These matter if you run long tasks, like coding sessions or personal assistants that need context. For one-off questions, you may never notice them.
Open weights and safety: what we know
Meta hasn’t said whether any Wonder model will be open-sourced. Alexandr Wang told Axios the company is “still committed to open source,” and Meta recently released weights for a separate model, Muse Glimmer.
On safety, Meta says it tightened safeguards against cyber and biological misuse. A model card is out, and a fuller report is promised later. Read that report when it lands before you trust the model with sensitive work.
How to test Muse Wonder before you commit
The fairest way to judge any new model is your own work. Pick a handful of real tasks you already run on another model, and compare the results side by side.
A simple test plan looks like this:
- Choose five to ten prompts from your daily work, including a few hard ones.
- Run them on Muse Wonder and your current model with the same settings.
- Track cost per task, not just price per token, since models use tokens differently.
- Check how well it remembers earlier context if you build assistants.
This takes an afternoon and tells you more than any launch chart. If Muse Wonder wins on your tasks, the switch is easy to justify.
Frequently Asked Questions
What is Muse Wonder?
It’s Meta’s first frontier-level AI model, released October 12, 2026. It’s built for coding, memory, and agent tasks.
How much does Muse Wonder cost?
Through the Meta Model API it costs $4 per million input tokens and $20 per million output tokens, per TechCrunch.
Can I use Muse Wonder Full?
Only if you subscribe to Muse Code. Meta doesn’t offer it through the API.
Is Muse Wonder better than Claude or GPT?
Meta says it’s comparable, but that comes from Meta’s own tests. Independent results so far are more mixed.
My take: give Muse Wonder a small test if you build with AI, but don’t move your main workflow yet. The API model costs the same as a stronger rival, and the best benchmark claims come from Meta itself. Wait for independent reviews and the full safety report.
Leave a Reply