Meta has introduced two new AI models, Muse Wonder and Muse Wonder Full. The company calls them its most capable models so far and says they move toward what it describes as personal superintelligence. The announcement landed Monday. Meta is trying to place the models against offerings from OpenAI and Anthropic, and it is making big performance claims in the process.
What Meta says the models can do
Meta says the pair adds Active Memory and Active Focus. Active Memory is meant to carry useful context from one conversation to another. Active Focus is designed to help an agent stay on difficult tasks for longer. Those are practical features, if they work as described. They target a common complaint about AI assistants: they forget what you told them and lose the thread on long jobs.
Benchmark scores draw quick pushback
Artificial Analysis, an independent evaluation platform, scored Muse Wonder Full at 56 on its Intelligence Index. The standard Muse Wonder scored 52. Both were tested at maximum reasoning effort, according to the article. That puts Wonder Full above Claude Fable 5.1 and GPT-6 Astra, which each scored 53 in the same index. It sits below Claude Opus 5.5, which scored 58.
Not everyone accepts the result. Bindu Reddy, CEO of Abacus.AI, which runs its own model evaluations, challenged Meta’s performance claims. She wrote on X that Muse Wonder was nowhere near Fable or Astra level. Reddy said detailed benchmark results would follow. That kind of dispute is common when new models arrive, especially when companies release headline numbers before independent testing settles.
Pricing and access
Meta’s new models are also its most expensive to date. The standard Wonder model charges $4 per million input tokens. That matches Claude Opus 5.5, according to the available comparison, but undercuts GPT-6 Astra and Claude Fable 5.1. Those models charge $10 and $50 respectively.
Wonder Full is not open to everyone. It is available only to subscribers of Muse Code’s $50 a month power usage tier or its $200 a month maximum usage tier. That access split may matter for developers who want the highest scores without paying enterprise-level prices. Meta, OpenAI, and Anthropic are all pushing similar claims about reasoning and memory. The real test will come from users, not just benchmark tables.
