Meta Launches Muse Wonder AI Models With Persistent Memory, Premium Pricing

Meta Platforms introduced two flagship artificial intelligence models on Monday, positioning them as the company’s most advanced systems yet and a step toward what it calls “personal superintelligence.” The release marks Meta’s most aggressive push into the premium AI assistant market, directly challenging Anthropic and OpenAI on both capability claims and pricing.

The standard Muse Wonder and the more powerful Muse Wonder Full introduce two architectural features Meta brands as Active Memory and Active Focus. Active Memory allows the models to retain and recall relevant details across separate conversation sessions — a persistent context window that functions more like a long-term user profile than a traditional chat history. Active Focus is designed to maintain reasoning coherence over extended, multi-step tasks that typically cause current models to lose track of objectives or hallucinate.

Independent benchmarking platform Artificial Analysis evaluated both models at maximum reasoning effort. Muse Wonder Full scored 56 on the platform’s Intelligence Index, while the standard Wonder scored 52. For comparison, Anthropic’s Claude Opus 5.5 leads the index at 58, while Claude Fable 5.1 and OpenAI’s GPT-6 Astra each sit at 53. The scores suggest Wonder Full narrowly outperforms the latest mid-tier models from its chief rivals but remains behind the current frontier leader.

Those numbers drew immediate skepticism. Bindu Reddy, chief executive of Abacus.AI — a firm that conducts its own model evaluations — posted on X that Muse Wonder was “nowhere near Fable or Astra level” and promised detailed benchmark results to support the claim. Abacus.AI has previously published comparative analyses that diverged from headline benchmark leaderboards, citing differences in evaluation methodology and real-world task performance.

Meta’s pricing strategy signals confidence. The standard Wonder model charges $4 per million input tokens, matching Claude Opus 5.5 while undercutting GPT-6 Astra at $10 and Claude Fable 5.1 at $50. Output token pricing was not disclosed in the announcement. Wonder Full, however, is gated behind Muse Code’s top subscription tiers: $50 per month for “power usage” or $200 per month for “maximum usage,” placing it firmly in the professional developer and enterprise segment.

The tiered access model mirrors OpenAI’s approach with GPT-6 Astra, which reserves its highest compute allocations for ChatGPT Enterprise and API customers on dedicated capacity plans. Anthropic offers Opus 5.5 through its API and Claude Pro subscription without a separate “maximum” tier, though rate limits apply.

Active Memory represents a significant architectural bet. Most current large language models treat each session as stateless unless developers implement external retrieval-augmented generation systems. By baking cross-session persistence into the model layer, Meta aims to reduce the engineering burden for personalized assistants — but raises fresh privacy and data governance questions. The company has not yet published details on data retention policies, user deletion rights, or whether Active Memory data informs future model training.

Meta’s announcement comes two weeks after the company’s Connect developer conference, where executives emphasized AI agents that can operate across WhatsApp, Instagram, and Messenger. The Muse Wonder family appears designed to serve as the reasoning backbone for those consumer-facing agents, though no integration timeline was provided.

Abacus.AI’s promised benchmark analysis is expected to publish later this week.