Meta's Muse Wonder Models Rival GPT-6 and Claude Opus — Why the Closed-Source Pivot?

Meta's New Flagship Models: Muse Wonder and Muse Wonder Full

On October 12, Meta released its next-generation flagship models, Muse Wonder and Muse Wonder Full (internal codename "Watermelon"). The models demonstrate strong performance in specialized tasks such as programming, legal analysis, and finance, with benchmark scores approaching or exceeding Anthropic's Claude Opus 5.5, Fable 5.1, and OpenAI's GPT-6 Astra.

Key Capabilities

Active Memory and Active Focus: The models retain important context across conversations and persist toward goals when encountering obstacles, eliminating the need for users to repeatedly restate objectives.

Ultra-Long Context: Support for 1 million token inference context enables processing of large codebases, hundreds of pages of contracts, or hours of audio/video recordings.

Massive Training Scale: Training was completed at the 1+ gigawatt "Prometheus" supercluster in Ohio, representing a massive compute investment.

Access and Pricing

Muse Wonder: Available via Meta Model API and Muse Code. API pricing: $4 per million input tokens, $20 per million output tokens, $0.20 per million cached input tokens.

Muse Wonder Full: Higher-performance variant, exclusive to Muse Code subscribers at $50/month (Power Usage) or $200/month (Maximum Usage). No standalone API offered.

Neither model provides downloadable weights; both require paid access, marking Meta's departure from prioritizing open-source models.

Unlike OpenAI and Anthropic's pursuit of artificial general intelligence, CEO Mark Zuckerberg is betting on personal superintelligence — AI assistants that can truly handle daily tasks and professional work. The new models will be integrated into Meta's personal AI agent Muse , initially for subscribers.

Why Meta Shifted to Closed-Source

Meta's pivot from its long-standing open-source approach stems from multiple strategic and practical factors:

1. Urgent Need for Commercial Monetization

While the Llama open-source series built a vast ecosystem, merely releasing weights failed to generate stable revenue. Open-source price wars continually depressed token resale value, and Meta could not capture core value from developers' downstream wrappers. By closing the models, Meta monetizes directly via API calls and subscriptions: Muse Wonder API pricing matches Claude Opus 5.5, while Muse Wonder Full's $200/month tier targets high-value enterprise and professional users, competing directly for enterprise contracts with OpenAI and Anthropic. This also lays groundwork for the Muse Agent ecosystem, where Meta controls the agent entry point and can later take commissions on transactions in e-commerce, travel, and other scenarios.

2. Strategic Inflection in Technical Competition

In 2025, Llama 4 fell short of expectations, and Meta's technical lead in the open-source track was rapidly erased. The performance gap between open and closed models has narrowed to 6–12 months; releasing weights alone no longer secures a frontier position. After 14 months of massive investment, training on the 1 GW Prometheus cluster produced a flagship model matching Claude Opus 5.5, placing Meta at the true industry frontier for the first time. Keeping the model closed protects this exclusive advantage. Novel mechanisms like active memory and active focus, plus breakthrough performance in legal and coding tasks, would be quickly replicated by competitors if open-sourced, eroding the technical moat bought with billions in compute.

3. Rebuilding Ecosystem and Business Moats

Closed-source forces developers and enterprises to depend entirely on Meta's model services, preventing self-hosted deployments and giving Meta full pricing and service control. This avoids the open-source problem where users can migrate to alternative open models at any time. Additionally, the 2026 U.S. AI regulatory framework targets only closed-source frontier models, excluding open models. Paradoxically, the closed path grants Meta access to government and high-sensitivity enterprise sectors that demand data security and service stability guarantees. Previously, Muse Agent was banned by platforms like Amazon because AI agents bypass traditional ad and recommendation systems; Meta must fully control model capabilities via closed-source to defend its new entry point in this platform power struggle.