Meta's First Frontier Model Is Priced Like One
Muse Wonder costs $4 in and $20 out, 40 times the input rate Meta charged for Muse Spark a month ago, and the version that actually reaches the frontier is locked inside a $50 subscription.
In September, Meta's best API model cost ten cents per million input tokens. On October 12, it launched one that costs $4.
That is the clearest signal in the release of Muse Wonder and Muse Wonder Full, which Meta Superintelligence Labs is calling its first frontier models. For a year Meta competed on price and openness while OpenAI and Anthropic held the top of the leaderboards. Wonder is the first time Meta has asked to be paid like a frontier lab. The question is whether the model has earned it.
What shipped
Two checkpoints, two very different distribution paths.
Muse Wonder is on the Meta Model API and OpenRouter at $4 per million input tokens and $20 per million output, with cached input at $0.20, according to the OpenRouter listing as reported by Forkast. It takes text, images, video, audio and PDFs, and has a context window of roughly 1 million tokens. The weights are closed.
Muse Wonder Full is not on the API at all. It is available only to subscribers of Muse Code, Meta's coding product, on the $50-a-month "power usage" tier or the $200-a-month "maximum usage" tier, Engadget reported. Meta says both models will also power its Muse personal agent, starting with paid subscribers.
Both were trained at Prometheus, Meta's gigawatt-scale data center in Ohio, according to TechCrunch. The headline features are active memory, which Meta describes as carrying what matters to you across conversations, and active focus, which it says lets an agent "keep working toward your goal when a task gets hard."
"We put a tremendous amount of effort into making sure your model can remember what's important to you," research scientist Kavya Kopparapu said in Meta's launch video.
The claim, and the scoreboard
Meta's model card says Wonder and Wonder Full reach benchmark scores comparable to Fable 5.1 and GPT-6 Astra. That is Meta's claim, built mostly from Meta's own testing, and Forkast noted that no research paper or ai.meta.com blog post accompanied the launch.
The most useful independent check so far is the Artificial Analysis Intelligence Index. On its leaderboard, Muse Wonder Full scores 56, tied with Claude Sonnet 5.5 and ahead of Fable 5.1 and GPT-6 Astra at 53. Claude Opus 5.5 leads at 58.
So the parity claim holds, for Full. Meta really has put a model inside the top cluster, roughly 14 months after Alexandr Wang's reorganization, which is not nothing for a lab that spent 2025 being written off.
The API model is a different story. Muse Wonder scores 52, ranked ninth of 216 models. That is the same score as GPT-6.1 Sol. Artificial Analysis puts Wonder's cost to run each index task at $4.52, against $0.72 for Sol. Same score, about six times the cost.
The pricing problem
TechCrunch described Wonder's API rate as roughly equivalent to Opus 5.5. That is the uncomfortable comparison. At about the same price per token, developers can buy the model at the top of the Artificial Analysis board, or Meta's model six points below it.
Meta has a few real answers. Wonder is fast, at 115 output tokens per second on Artificial Analysis, though its time to first token was over 20 seconds there. On MemoryBench, Meta's own continual-learning benchmark, Meta says Wonder beats Fable, Opus and Astra. The benchmark is Meta's, and it uses an LLM judge, so it is a reason to test the model rather than a reason to switch.
And memory is plainly where Meta wants to compete. A model that remembers you across weeks is worth more to a company with billions of users than one that tops a coding leaderboard by two points.
Why gate the good one
Putting Wonder Full behind Muse Code instead of the API looks like a deliberate choice, not a capacity constraint.
Serving a frontier model through an API invites direct price comparison, token by token, with Anthropic and OpenAI. Bundling it into a $50 subscription turns it into a product. Subscribers compare the experience, not the per-token rate, and Meta captures a monthly relationship rather than a metered call that can be routed elsewhere tomorrow.
It also fits the money. Engadget notes Meta expects to spend up to $145 billion on AI this year. Mark Zuckerberg told the Financial Times that reaching this point "required a large amount of compute and investment, and a lot of very rigorous scientific work." Ten-cent tokens were a land grab. A $50 tier with the best model inside is a revenue line.
The trade-off falls on developers. The strongest Meta model is not something you can build on, only something you can use, and the one you can build on is priced against stronger competition.
What to watch
Independent memory evals. Meta has staked the launch on active memory. Until someone outside Meta tests recall over long, messy, multi-week use, MemoryBench is a claim.
Whether Full reaches the API. If Meta opens it at a price below Opus 5.5, the scoreboard math changes overnight.
Where Muse Code pricing lands. A $50 plan with a top-cluster model is a direct shot at the coding subscriptions OpenAI and Anthropic sell. If uptake is strong, expect Meta to keep its best checkpoints inside its own products and treat the API as the second shelf.