Meta's Muse Wonder reaches the AI frontier, but not the top

Meta Superintelligence Labs released Muse Wonder on October 12, the first model the company describes as frontier-class since Mark Zuckerberg created the unit last year. A week of independent testing now shows where it lands: within a point of OpenAI's GPT-6 Astra and Google's Gemini 4 Argon, and still behind Anthropic's two strongest models.

The company's own framing is bolder. Meta's model card claims "comparable benchmark scores to Fable 5.1 and GPT-6 Astra," and Chief AI Officer Alexandr Wang wrote on X that the larger sibling, Muse Wonder Full, "rivals the big bois." For a company whose last open flagship, Llama 4, landed with a thud, even parity is a reset.

What Meta is selling, and for how much

Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with a 1-million-token context window, according to The Futurum Group's breakdown. Both models are closed-weight, a clean break from the Llama strategy. Wonder Full will not be served through the API at all, TechCrunch reported; it is reserved for Muse Code subscribers on $50 or $200 monthly plans, according to CNBC. Meta trained both at its Prometheus data center in Ohio.

Meta has pitched the API as cheaper than OpenAI and Anthropic. That holds against Fable 5.1 and Astra, which list at $10 and $50 per million tokens, but not against Anthropic's Opus 5.5, which Futurum says carries the same $4/$20 rate card.

Independent scores put Wonder at 52, not first

Artificial Analysis, which runs its own evaluations, gives Wonder 52 on its Intelligence Index and Wonder Full 56, up from 48 for Muse Spark 1.3, its predecessor. On the same leaderboard, Opus 5.5 leads at 58 and Sonnet 5.5 sits at 56, while Fable 5.1, GPT-6 Astra and Gemini 4 Argon cluster at 53. Wonder Full is effectively tied with Sonnet; Wonder sits a point below the trio it says it matches.

Where Wonder does stand out independently is restraint. Artificial Analysis measured a 27% hallucination rate for Wonder and 23% for Wonder Full. Meta's more striking numbers, such as 85.6% on MemoryBench and 53.1% on SWE-Marathon against 46.3% for Opus 5.5, are company-reported and were run on different test harnesses, Futurum found.

"Every figure Meta published was run at maximum effort, however, and most of its agent results come from Meta's own evaluations, so buyers should treat them as an upper bound," said Nick Patience, who leads AI platforms research at Futurum. He also noted that Meta left Gemini 4 Argon, arguably its closest rival, out of the launch comparisons.

Why Meta needed a frontier model now

Zuckerberg has said he was "not happy with the trajectory that we were on with Llama 4," Fortune reported. Released in April 2025, Llama 4's public Maverick model ranked below rivals on LMArena after Meta had promoted an experimental variant on the leaderboard. That June, Zuckerberg created Meta Superintelligence Labs, having paid $14.3 billion for 49% of Scale AI to bring Wang in to run it.

The bill since has been enormous. Meta expects 2026 capital spending of $130 billion to $145 billion, according to Fortune, and Wonder is the first real test of whether that money can produce revenue outside advertising. CNBC framed the paid Muse plans as a check on users' willingness to pay. WIRED's Will Knight called the release "a significant step forward for Meta's models," while cautioning that benchmarks are not always a reliable measure of real-world performance.

Next comes a commercial test rather than a technical one: whether developers move API traffic to Wonder when Opus 5.5 costs the same and scores higher. Patience put the risk plainly: "a strong model that is hard to buy will struggle in a market this crowded."