Meta Superintelligence Labs released two new models on October 12, Muse Wonder and Muse Wonder Full, which TechCrunch and other outlets describe as the company's first frontier-class model. The release splits access sharply: the standard Muse Wonder is available through the Meta Model API and Muse Code, while the stronger Muse Wonder Full is restricted to Muse Code subscribers and is not sold per token.

Access and pricing

According to launch-week coverage, Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with a 1M-token context window and discounted cached input. Muse Wonder Full has no public per-token price. Reports say it is included in Muse Code's paid tiers, listed at $50 and $200 per month. Outlets differ slightly on how those tiers are named, so check Meta's pricing page before relying on plan labels.

Meta highlights two traits it says come from how the models were trained rather than from settings users switch on. Active Memory is meant to retain the details that matter across long conversations and bring them back when relevant. Active Focus is meant to keep the model working on difficult goals and to check in before important actions.

How it benchmarks

Third-party and Meta-reported figures should be read separately. On the Artificial Analysis Intelligence Index, reports put Muse Wonder Full at 56 and Muse Wonder at 52, up from 48 for Meta's earlier Muse Spark 1.3. The same reports list GPT-6 Astra at 53 and Claude Opus 5.5 at 58, which places Wonder Full between them. Coverage also cites coding results such as SWE-Marathon and DeepSWE 1.1 and a memory benchmark where Muse Wonder scored ahead of several rivals; many of these numbers come from Meta's own testing and have not been independently reproduced.

There are caveats. Artificial Analysis reportedly saw heavy token use, about 130 million output tokens to run its index, which can erode a per-token price advantage. Early testers also reported hitting usage limits quickly on the $200 Muse Code plan.

Why it matters

Meta had been a step behind the leading labs at the frontier, and Wonder Full puts it in the same conversation on at least one widely followed index. The decision to hold the stronger model back from the API is the notable strategic choice: developers who want the best Meta model must adopt Meta's own coding product, while the API gets the second-best variant. For teams evaluating agents built on Meta's models, the practical question is whether the benchmark gains hold up in independent testing and whether token consumption keeps real-world costs competitive.