Meta pushes new Muse Wonder models

Meta unveiled two AI models, Muse Wonder and Muse Wonder Full. The company says they are its most capable models to date and a step toward what it calls personal superintelligence. That is not a small claim. The launch puts Meta more directly against OpenAI and Anthropic in the high-end model market.

Active Memory and Active Focus

The models introduce two features Meta calls Active Memory and Active Focus. Active Memory is meant to carry relevant information across conversations. So a user does not have to repeat context each time. Active Focus helps the agent stay on difficult tasks. That could matter for coding, research, or long multi-step work. Meta says these are central to how the models handle extended sessions. Maybe the real test is whether they work reliably outside demos.

Benchmark scores draw attention and doubt

Artificial Analysis scored Wonder Full at 56 on its Intelligence Index. Standard Wonder scored 52. Both scores were at maximum reasoning effort. That places Wonder Full above Claude Fable 5.1 and GPT-6 Astra. Each of those scored 53. It still trails Claude Opus 5.5, which scored 58. The gap is not huge, but it is enough to question Meta’s framing. Bindu Reddy, CEO of Abacus.AI, challenged the performance claims. She wrote on X that Muse Wonder was nowhere near Fable or Astra level. She said detailed benchmark results would follow. Independent evaluations often differ, so the final picture may take time.

Pricing and access

Meta’s new models are its most expensive to date. Standard Wonder costs $4 per million input tokens. That matches Claude Opus 5.5. It undercuts GPT-6 Astra and Claude Fable 5.1, which 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 or $200 a month maximum usage tiers. That makes Full a premium product for heavy users. For now, Meta is making a strong claim with a mixed scorecard. It has one model ahead of some rivals on one index, but behind the top score. And it faces public doubt from another evaluator. The next few weeks should show whether developers agree. Or not. Benchmark numbers are useful, but they rarely tell the whole story.