Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API

Laura Meta did the annoying thing again, Harper. The stronger model is the one you can’t actually buy through the API, which is such a perfectly Meta move that I almost respect it.

Harper Yeah, that’s the whole story right there. They’re saying frontier, but they’re also saying, very quietly, that the frontier is inside their own product unless you’re willing to take the lesser one.

Laura And that matters because the model that ships is the model that changes behavior. If the only externally usable thing is Muse Wonder, then enterprise teams are evaluating a middle layer, not the thing Meta is bragging about.

Harper Right. The article’s benchmark table is the tell. Wonder Full ties Claude Sonnet 5.5 on Artificial Analysis, sits ahead of GPT-6 Astra and Gemini 4 Argon there, but Meta is still withholding it from API buyers.

Laura Mm-hm.

Harper So the mechanism isn’t mysterious. Meta trained both on Prometheus, used big initial training plus reinforcement learning, then split distribution by product surface. That’s not a science story first. It’s a control story.

Laura Which is why the user story gets weird fast. If you’re a team looking for a serious agent model, you care about access, pricing, and whether your ops people can tolerate it. Not just whether some leaderboard went green for a day.

Harper Exactly. And the model card is doing a little too much work. Meta mixes internal evals, vendor-supplied scores, and third-party results. That doesn’t make the claim fake, but it does mean you should trust the independent Artificial Analysis numbers more than the glossy launch framing.

Laura Okay, but the independent numbers are still pretty good for them. Wonder Full scores fifty-six on that index, tied with Sonnet 5.5, and the standard Wonder lands at fifty-two. That is not a toy release.

Harper No, it’s not. The part that jumps out to me is the sustained-work stuff. On SWE-Marathon, Wonder Full is fifty-seven point five versus forty-six point three for Opus 5.5. On the legal benchmark it’s thirteen point one, which sounds huge until you remember everyone is still kind of bad at legal deliverables.

Laura That’s the useful nuance, though. If you’re buying this for long-running coding or agent loops, you probably care more about that kind of spread than about a single aggregate score. It’s the difference between a model that looks good in a demo and one that survives a messy task.

Harper Right, right.

Laura And Meta is leaning hard into the product framing here. Active Memory and Active Focus are basically the pitch that the model remembers relevant context and keeps going when it hits a wall. That’s the part enterprise buyers actually understand.

Harper I buy the framing more than the hype, but only partly. Access to prior details and plan persistence are real capabilities, sure. The low bar is that systems have been tracking state forever. The hard part is doing it without wandering off or becoming confidently wrong for twenty minutes.

Laura Oh, that is such an Exploring Next sentence.

Harper I know. It hurts because it’s true.

Laura And honestly, the pricing tells you who Meta thinks this is for. Four dollars input, twenty dollars output per million tokens is premium territory. That’s not a bargain open model play. That’s a serious workhorse price.

Harper Yeah, and the comparison is telling. It undercuts GPT-6 Astra and Claude Fable 5.1 on list rate, but sits right next to Claude Opus 5.5. So Meta isn’t trying to win on cheapness. They’re trying to look like a peer in the expensive tier.

Laura Mm-hm.

Harper The catch is cost per task. Artificial Analysis says an average Intelligence Index task costs four dollars and fifty-two cents on Wonder, versus five dollars and ninety-eight cents on Opus 5.5 and three dollars and twenty-six cents on GPT-6 Astra. So token price and task price are already diverging.

Laura Which is exactly why I care about this. Buyers do not buy tokens. They buy outcomes. If Wonder is pricier on the meter but cheaper on the completed workflow, that is the actual enterprise question.

Harper And that’s where Meta still has to prove it outside its own walls. The article points out the API supports OpenAI- and Anthropic-compatible tooling, which lowers friction. But the strongest model living only in Muse Code means the best case is still partly captive.

Laura I’d put a bet on the obvious part: by the end of the month, everyone talking about Muse Wonder will be talking about the gate, not the benchmark. The strongest model inside the product becomes the headline, not the footnote.

Harper I’ll take the other side a little. I think the benchmark spread will keep the conversation alive longer than that, especially the long-running agent numbers. But I’d still bet the API story is what enterprise teams notice first.

Laura Yeah, because that’s the thing that ships. Not the pretty slide. The thing they can route through procurement without a side quest.

Harper Exactly. And if Meta really wants this to matter beyond a launch cycle, they’ll have to decide whether Wonder Full is a premium teaser or an actual platform move.

Laura Okay, that’s the part I’m watching too. And I cannot believe this is what we’re doing on a Tuesday, but here we are.

Harper Exploring Next, baby. Our finest hour.

Laura Fine, your finest hour. I’m taking the win on the product gate, and I’ll let you keep the benchmark spreadsheet. Talk tomorrow.