Meta’s Best New Model Isn’t on the API. It’s Inside the Subscription That Trains on Your Code.

TL;DR: On 12 October Meta released its first frontier models: Muse Wonder and Muse Wonder Full. Only the weaker one is on the API. The better one is available solely inside a Muse Code subscription — and Meta’s own help page says those subscriptions default to Contributor mode, where your prompts, code and files are used to train future Meta models. In August we said you could simply pay the standard rate to avoid that. That escape route no longer reaches Meta’s best model.

What Meta shipped

TechCrunch reported on 12 October that Meta’s Superintelligence Lab released two models — Muse Wonder and Muse Wonder Full — described as the company’s first genuine attempt at a frontier model. Meta’s own blog called them “a major leap toward personal superintelligence”. Both Muse Wonder models are much larger than Muse Spark, the lightweight model we wrote about in August.

Meta is leaning hard on memory. Research scientist Kavya Kopparapu said in the launch video: “We put a tremendous amount of effort into making sure your model can remember what’s important to you.” Meta claims Muse Wonder beats Fable, Opus and Astra on MemoryBench, a continual-learning benchmark.

VentureBeat’s write-up is the one worth reading, because it puts the split in the headline. Neither model has downloadable weights. Both take text, image, audio and video in, and both carry a context window of roughly one million tokens.

The Muse Wonder split, and why it matters

Here’s the shape of it:

	Muse Wonder	Muse Wonder Full
Where you get it	Meta Model API (`muse-wonder-1.0`)	Muse Code subscription only
Price	$4 per million input tokens, $20 per million output	$50/month (Power Usage) or $200/month (Maximum Usage)
Pay-as-you-go API	Yes	None announced
Data default	Standard — Meta says your content isn’t used for training	Contributor — your content is used for training

That bottom row is the whole story. The two paths to Meta’s new models have opposite data defaults, and the better model is only on the path that defaults to training on you.

The independent benchmark, not Meta’s

Most of the numbers in Meta’s launch materials are Meta’s own. One isn’t: the Artificial Analysis Intelligence Index, which VentureBeat reports as follows.

Model	Index score
Claude Opus 5.5	58
Muse Wonder Full	56
Claude Sonnet 5.5	56
GPT-6 Astra · Gemini 4 Argon · Claude Fable 5.1	53
Muse Wonder	52
Grok 4.6	44

Read that carefully, because it is the opposite of the headline. The model you can actually build on sits at the bottom of the frontier pack — below GPT-6 Astra and Fable 5.1. The one that nearly tops the table is the one you can’t get without a subscription.

VentureBeat is also sharp about Meta’s own figures: the comparisons mix internal, vendor-supplied and third-party results, some coding tests use different tooling, and Meta trails Opus and Astra on Terminal-Bench even by its own numbers. Treat the coding claims as a vendor’s claims.

What changed since August

We covered this trade when Muse Code launched: Meta’s coding AI was 21x cheaper if you let it train on your code. Back then the structure was two API prices sitting side by side, and we said — fairly, we think — that Meta deserved credit for putting the choice in the open, and that anyone doing client work should simply pay the standard rate.

Two things have moved, and both are worse for the person deciding.

First, it’s a default now, not a menu. Meta’s own help page states it plainly: “Muse Code subscriptions default to Contributor mode.” What that covers is broad — “the prompts, code, files, and context you send, plus the model responses you get back.” You can switch it off in settings at any time, but the page is explicit that this stops future content being used. It says nothing about what has already gone.

Second, paying more no longer buys you out. In August, the standard rate got you the same model without the training. Today the top model isn’t on the standard path at all. The subscription tiers run $5, $15, $50 and $200 a month, and Muse Wonder Full starts at the $50 Power Usage tier.

So the August advice — pay the standard rate — still protects your code, but it now costs you the capability as well. That’s a different bargain, and we’d rather say so than pretend our earlier read still covers it.

Does this actually affect you?

Honestly: for most people reading this, not yet.

Muse Code is a terminal coding agent, which already puts it outside what most non-developers will touch, and nothing about this release changes that. If you’re building with Claude Code, Cursor, Lovable or Replit, nothing switched today and there’s no action to take.

What you should take from it is the pattern, because it’s the second time this year we’ve watched it tighten:

1. A capable model arrives with two access routes.
2. The routes have different data terms.
3. Over time, the better capability drifts towards the route with the looser terms.

That is a thing to recognise in any tool, not just Meta’s. The question to carry around is no longer “does this tool train on my data?” but “what does it cost me to say no, and is the good version still available if I do?“

What to do

1. If you don’t use Muse Code, do nothing. This is a pattern to file, not a product to adopt.
2. If you do use it, open the settings and check Contributor mode. It’s on unless you turned it off, and turning it off only affects what you send from that point.
3. Decide the question once, in advance. Would you be comfortable with this code, these files and these prompts training a model? Hobby project — probably fine. Client work or the thing that makes your business yours — not fine, at any discount.
4. Don’t buy a model on a vendor’s benchmark table. Meta’s coding numbers are Meta’s. The one independent index puts the Muse Wonder you can actually buy mid-pack.
5. Watch who copies this. We flagged the three-labs pattern when GPT-6 Astra landed; capability-behind-looser-terms is the version worth watching now.

Who should care (and who shouldn’t)

• Using Muse Code today: check the Contributor setting this week. That’s the whole job.
• Building a business on any AI coding tool: read your tool’s data terms once, properly, and write down the answer. You will be asked this by a client eventually.
• Building with hosted app builders: unaffected — you’re not sending code to a model API directly, and your builder’s terms are the ones that matter.
• Shopping for a model by benchmark: this release is a good lesson in why the asterisk matters more than the headline score.
• Everyone else: nothing to do. File the pattern.

Our take

The split is the news, and Meta hasn’t explained it.

There’s a reasonable commercial logic to keeping your best model inside your own product — plenty of vendors do it. But pairing that with a subscription that defaults to training on the user’s code means the people getting the best capability are the ones contributing the most training data, whether they noticed or not. Those two decisions are defensible separately and uncomfortable together.

We’d revise our August line slightly. We praised Meta for putting the trade in the open, and that was right about the API, where two prices sat side by side and you picked one. A default inside a subscription is a weaker kind of consent than a price you deliberately choose, even when it’s documented — and it’s documented clearly, to be fair.

On the model itself: unconvinced for now. A 52 on the one index nobody at Meta controls is not a frontier result, whatever the model card says, and the 56 is locked behind $50 a month. If you want the best coding model today, this isn’t the reason to move.

The thing we’d actually watch is whether the API catches up. If Muse Wonder Full reaches the standard-terms API in a few months, this was a launch sequencing story and nothing more. If it doesn’t, Meta has quietly made “train on my code” the price of its best work, and that is a precedent other vendors will notice.

Not sure what your current tool does with your code? Take the 60-second Vibe Coding Tool Finder quiz →

FAQ