Meta Releases Muse Wonder Frontier AI Models to Compete With OpenAI and Anthropic

12 Sources

Share

Meta unveiled Muse Wonder and Muse Wonder Full, its first frontier AI models from the restructured Superintelligence Labs. The models rival GPT-6 Astra and Claude Opus 5.5 with advanced active memory capabilities. Access requires paid subscriptions starting at $20 per month for personal agents.

Meta Returns to Frontier AI Competition With Muse Wonder Launch

Meta released two state-of-the-art models on Monday—Muse Wonder and Muse Wonder Full—marking the company's first attempt at frontier AI models

1

2

. These releases represent a significant milestone for Meta after Mark Zuckerberg restructured the company's AI efforts last year, sinking billions into the newly formed Meta Superintelligence Labs under the leadership of former Scale AI CEO Alexandr Wang

2

. According to Meta's internal testing, Muse Wonder achieves benchmark scores comparable to OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1

1

. The company says Muse Wonder Full even outperforms Claude Fable 5.1 and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index

2

. Meta trained these models at its Prometheus data center in central Ohio, which came online earlier this year

1

.

Source: Wired

Source: Wired

Paid Subscription Model Signals Revenue Diversification Strategy

Access to Muse Wonder requires paid subscriptions, with Meta positioning the models as premium offerings for digital assistants

4

. Users of the Muse AI agent app must subscribe to monthly plans priced at $20 or $100 to access the new capabilities

4

. Through the Meta Model API, Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens, making it roughly equivalent in cost to GPT-6 Sol or Claude Opus 5.5

1

. Muse Wonder Full will only be available to Muse Code subscribers under $50 or $200 monthly subscription tiers

4

5

. This pricing strategy underscores Meta's efforts to create AI-derived revenue sources outside its core online advertising business

4

.

Source: Interesting Engineering

Source: Interesting Engineering

Active Memory Feature Targets Long-Term Agent Performance

Meta emphasized active memory as a breakthrough capability that allows Muse Wonder to retain and synthesize information over extended periods

1

5

. Research scientist Kavya Kopparapu stated that Meta invested tremendous effort into ensuring the model remembers what matters to users over days, weeks, months and beyond

1

. The company highlighted its performance on MemoryBench, a benchmark challenging continual learning and memory construction frameworks, where Muse Wonder now scores higher than Claude Fable, Claude Opus, and GPT-6 Astra

1

. Alongside active memory, Meta introduced active focus to keep agents locked into stated goals and prevent confusion during complex tasks like research or large software projects

2

5

. Alexandr Wang explained that the model was specifically trained to try alternative approaches when encountering roadblocks instead of stalling and checking back with users

4

.

Specialized Performance in Coding and Professional Tasks

The Muse Wonder models demonstrate frontier coding performance and excel at navigating legal, financial, and health-related tasks

5

. Meta's testing shows the models outperform GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 on specialized benchmarks measuring legal document analysis and software development capabilities

2

. Both models will power the Muse AI agent, with Meta planning to develop specialized variants for broader consumer deployment, though the company hasn't shared specific timelines

3

. The models represent a major advancement from Muse Spark, Meta's lightweight model released in April, which prioritized low latency over frontier capabilities

1

2

.

Source: ET

Source: ET

Safety Measures and Agent Monitoring Amid Industry Concerns

Meta claims Muse Wonder offers the strongest safeguards to date against cyber misuse, with significant training focused on refusing dangerous outputs like bioweapon recipes or hacking assistance

2

4

. Mark Zuckerberg stated last month that Meta delayed the model release to work on safety improvements

2

. The company revealed a framework designed to scale model capabilities safely in a blog post earlier this month

2

. Wang said Meta's AI team monitors for agent-related incidents with the Muse app, addressing user complaints about unauthorized message access by educating users on authorization processes

4

. However, Meta didn't disclose whether it observed misbehavior during testing or specific mitigation strategies, despite recent incidents involving rogue AI agents from Anthropic and OpenAI

2

.

Implications for Meta's AI Ambitions and Industry Competition

The Muse Wonder launch validates Meta's massive investment in AI infrastructure, with the company expecting to spend as much as $145 billion this year on AI initiatives

5

. Zuckerberg told the Financial Times that reaching this milestone required both enormous compute resources and rigorous scientific work from the superintelligence team

5

. Meta had fallen behind after delivering disappointing Llama model versions while Anthropic, OpenAI, and Google rapidly introduced more capable systems

2

. The company's path toward personal superintelligence now appears more credible, though questions remain about how Muse Wonder compares to Google's Gemini 4 Argon, whose capabilities are still emerging

3

. Watch for Meta's planned open-weight model releases and how effectively active memory translates to real-world agent performance in booking reservations, managing schedules, and handling complex professional tasks over extended timeframes.

Today's Top Stories

© 2026 TheOutpost.AI All rights reserved