Meta's Muse Wonder Is Frontier AI at Opus Prices — But the Best Version Is Subscription-Only

Meta has finally shipped a frontier model. The company says so itself — for the first time, Meta Superintelligence Labs is calling its new models frontier-class, not just competitive. And with Meta Muse Wonder, the headline move isn't the benchmarks. It's the price list.

Meta Muse Wonder costs exactly what Claude Opus 5.5 costs: $4 per million input tokens, $20 per million output, $0.20 for cached input. Down to the cent. That's not a discount strategy — that's a declaration that Meta considers itself a peer of Anthropic now, and it wants the invoice to prove it.

There are two models in the family: Muse Wonder and Muse Wonder Full. Wonder is the one on the Meta Model API. On the Artificial Analysis Intelligence Index, Wonder scores 52 and Wonder Full scores 56, against Opus 5.5's 58. Close. But here's the thing the launch materials bury a little: Wonder Full, the variant that does most of the winning in Meta's own benchmark table, is not on the API at all. It lives inside Muse Code, behind the $50-per-month Power Usage and $200-per-month Maximum Usage plans.

Meta Muse Wonder pricing: matching Opus 5.5 to the cent

The API price parity is worth pausing on, because pricing in this market is positioning. Meta could have undercut Opus 5.5 by half and flooded developers toward its own stack — that's the classic Meta playbook. Instead it matched the price exactly, which reads like confidence: we think our model is as good as the reference model, pay us the same. Whether the market agrees will show up in routing data over the next few months.

The economics aren't trivial either. Artificial Analysis puts the cost per index task at roughly $4.52 for Wonder versus about $1.60 for Meta's previous Muse Spark 1.3 — nearly 3x the cost of running its own last-gen model. Frontier pricing is frontier pricing; the cheap route is still the older models, as the AI agent wars keep demonstrating.

What's actually new: Active Memory and Active Focus

The genuinely interesting part of the launch is the feature pair Meta is betting on. Active Memory lets the model retain key facts across conversations, for days to months — not as a context-window trick, but as persistent memory. Active Focus is the agentic side: the model sticks with hard goals, tries alternate approaches when it gets stuck, and checks in before taking big actions.

If you've been reading about how AI agents actually work in 2026, you'll recognize the pattern: the frontier is no longer raw benchmark scores, it's behavior over long horizons. Memory that persists and focus that survives failure are exactly the capabilities that separate a useful coding agent from a clever autocomplete. The 1M-token context window, multimodal inputs (text, images, video, audio), and text output are table stakes at this point.

The benchmark fine print

Meta's own table shows the Wonder family leading 7 of 14 benchmark rows to Opus 5.5's 6 and GPT-6 Astra's 1. Sounds like a win — until you notice that most of the Wonder wins come from the Full variant, the one you can't buy via API. Wonder, the model developers actually get, is a notch behind across the board.

The weak spot worth naming: factual knowledge. Wonder scores just 30 on the AA-Omniscience test, below rivals. For a model pitched at coding agents and long-horizon work, factual wobbles are the failure mode that bites — confident, wrong answers at the end of a 40-step task.

The privacy wrinkle nobody is talking about

Buried in Meta's model page: Muse Code trains on user code by default unless you opt out. For a company trying to win developers from GitHub Copilot and Cursor, this is a live grenade. Enterprise teams that already balk at Anthropic or OpenAI data policies are not going to find Meta's default more comforting. If you're an admin evaluating this — and Google's own Argon just became an option in Copilot — read the data policy before the benchmark table.

What Meta is really signaling

Trained at the gigawatt-scale Prometheus data center, announced by Alexandr Wang on X, with a launch-day HN thread past 600 points: this is Meta's coming-out party as a frontier lab. The price matching is the strategy document. Meta isn't racing to the bottom on inference costs; it's racing to be taken seriously as the lab that sets prices rather than undercuts them.

Whether that confidence is justified depends on the next three months of independent evals and developer adoption. But one thing is already clear: the AI tool stack conversation just got a new contender at the top tier, and it's not the cheapest option in the room.