Muse Wonder is Meta’s new frontier model for AI agents, and you can use it today through the Meta Model API or the Muse Code coding tool. Meta announced it on October 12, 2026, along with a bigger sibling, Muse Wonder Full, that sits behind a paid plan.
Here is the short version before we get into the details:
- Announced October 12, 2026 by Meta.
- Muse Wonder is open to developers through the Meta Model API and Muse Code.
- Muse Wonder Full is only for paid Muse Code subscribers.
- API price: $4 per million input tokens and $20 per million output tokens, per MarkTechPost.
- Both models handle a 1 million token context window.
That list is the whole story at a glance. The rest of this post covers what each model does, what it costs, and whether the benchmark claims hold up.
What Muse Wonder and Muse Wonder Full actually are
Meta calls both models “frontier” and built them for agentic work, meaning tasks where the AI plans, tries things, and fixes its own mistakes. According to Meta’s announcement, the focus is coding, plus expert areas like law, finance, and health.
Two features get most of the attention. Active Memory carries useful details from one conversation to the next. Active Focus keeps the model working toward a goal when a task gets hard, so it tries another approach, recovers from errors, and checks in before anything important.
Meta also says the models are trained not to share private information without permission, and to still help with legitimate security work like finding flaws in your own code. A full safety report is promised “soon,” so treat the safety claims as unverified for now.
Muse Wonder will also start powering the regular Muse assistant, first for Muse subscribers. Meta has not given a date for that.
What Muse Wonder costs and who gets which model
Meta’s own announcement does not list prices. The numbers below come from MarkTechPost’s rundown of Meta’s developer docs, so check dev.meta.ai before you budget around them.
| Model | Access | Price |
|---|---|---|
| Muse Wonder | Meta Model API and Muse Code | $4 input, $0.20 cached input, $20 output per 1M tokens |
| Muse Wonder Full | Paid Muse Code plans only | Included in Power ($50/month) or Maximum ($200/month) |
Both models take text, images, video, and audio as input and reply with text. They run only on Meta’s cloud, so there is no downloadable version. If you want open weights, Mistral Large 4 and Kimi K3.1 Pro are the closer fit.
The API price matches Claude Opus 5.5 at $4 and $20, and it is well below Claude Fable 5.1 and GPT-6 Astra at $10 and $50. Cheap per token does not always mean cheap per task, though, since agents can burn a lot of tokens retrying.
How Muse Wonder stacks up on benchmarks
Meta published its own comparison table, which MarkTechPost reproduced. Higher is better in every row, and these are Meta’s numbers at maximum reasoning effort.
| Benchmark | Wonder Full | Wonder | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| DeepSWE 1.1 | 79.0 | 78.8 | 74.2 | 74.1 |
| SWE-Marathon | 57.5 | 53.1 | 46.3 | 42.5 |
| ProgramBench | 59.0 | 50.0 | 70.5 | 52.5 |
| OSWorld V2.1 (partial) | 78.4 | 72.6 | 81.8 | 75.8 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 59.6 | 59.1 |
The pattern is mixed. Muse Wonder Full leads on long software-engineering tasks, while Claude Opus 5.5 wins on ProgramBench, computer-use tasks, and the terminal test. On the independent Artificial Analysis index, Opus 5.5 scores 58 against 56 for Wonder Full.
Remember that these are vendor-run results. Meta says its competitor runs are best-effort and may not be tuned, and some of its tests, like MemoryBench, are internal. Wait for independent testing before you switch a production workload.
How to try Muse Wonder today
There are two routes, depending on whether you want an API or a coding assistant in your terminal. Per MarkTechPost’s walkthrough, the API route goes like this:
- Sign in at dev.meta.ai with a Meta account.
- Add a payment method under Billing, because the “Create API key” button stays disabled until you do.
- Create a key and set it as the
MODEL_API_KEYenvironment variable. - Point your OpenAI-compatible client at
https://api.meta.ai/v1and use the model namemuse-wonder-1.0.
That compatibility means many existing tools can switch over by changing a base URL and a key.
For Muse Code, install it with the script on dev.meta.ai (it runs on macOS, Linux, and Windows), then run muse in your project folder. The default model is muse-spark-1.2. Switch with /models or start it with muse --model muse-wonder-1.0.
Frequently Asked Questions
Is Muse Wonder free?
No. The API is pay-as-you-go and needs a payment method on file. Muse Wonder Full needs a paid Muse Code plan.
What is the difference between Muse Wonder and Muse Wonder Full?
Wonder Full is the stronger version and scores higher on most of Meta’s tests. It is only available inside paid Muse Code plans, while Wonder is also sold through the API.
Can I run Muse Wonder on my own computer?
No. Both models run only on Meta’s cloud, and Meta has not announced open weights.
Is Muse Wonder better than Claude or GPT-6?
It depends on the task. Meta’s numbers show it ahead on some coding tests and behind Claude Opus 5.5 on others, and none of it is independently confirmed yet.
Is Muse Wonder worth trying?
Yes, if you already pay for a coding assistant and want a cheap second opinion. At $4 and $20 per million tokens with a 1M context window, a side-by-side test on your own codebase costs very little. I would not move a production workflow over until third parties confirm the benchmark claims and Meta ships that safety report.
Leave a Reply