Meta announces its first frontier model, Muse Wonder

Summary

Meta's Superintelligence Lab has unveiled its first frontier-class models, Muse Wonder and Muse Wonder Full. The company claims these models perform on par with Fable 5.1 and GPT-6 Astra in benchmarks and lead in its proprietary MemoryBench metric. This launch marks Meta's high-level model development initiative following Alexandr Wang's restructuring of its AI efforts in August 2025.

Why it matters

This move signals that Meta has finally entered the frontier-class competition, narrowing the gap between OpenAI and Anthropic in a market where model capability drives enterprise adoption. Pricing ($4/$20 per million tokens) positions the product directly against GPT-6 Sol and Opus 5.5, posing a significant commercial threat. Furthermore, the emphasis on long-term 'active memory' suggests Meta views the next product frontier as agentic and persistent assistants rather than one-off chats.

Highlights

• Self-reported benchmarks claim performance parity with Fable 5.1 and GPT-6 Astra, along with a lead over Fable, Opus, and Astra on MemoryBench
• Wonder Full is locked behind Muse Code subscriptions and inaccessible via public APIs, creating a two-tier access model
• The models were trained at Meta's Prometheus data center in Ohio, which came online in early 2026

Critical take

All performance claims rely on Meta's own model card and proprietary MemoryBench metric, lacking independent third-party evaluations. The term 'comparable' conveniently avoids clarifying whether Wonder actually outperforms or merely matches current rivals. Additionally, locking the more powerful Wonder Full behind a paid subscription tier while marketing the pair as 'frontier-class' raises questions about how much of this claimed capability is truly accessible to the broader developer ecosystem.