Meta Unveils Muse Wonder Frontier AI Models, Claims Performance Matching OpenAI and Anthropic

Reviewed byNidhi Govil

46 Sources

Share

Meta released Muse Wonder and Muse Wonder Full, marking its entry into frontier AI models after a 14-month overhaul. The models claim benchmark scores comparable to GPT-6 Astra and Claude Opus 5.5, with pricing at $4 per million input tokens. Meta emphasizes personal superintelligence over knowledge work automation.

Meta Returns to Frontier AI Competition With Muse Wonder Launch

Meta's Superintelligence Labs released Muse Wonder and Muse Wonder Full on Monday, representing the company's first frontier AI models since Mark Zuckerberg restructured its AI efforts 14 months ago

1

2

. These models mark a decisive shift from Meta's lightweight Muse Spark model toward competing directly with OpenAI and Anthropic at the cutting edge of AI development. According to Meta's internal testing, Muse Wonder achieves benchmark scores comparable to GPT-6 Astra and Claude Fable 5.1, while Muse Wonder Full ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index

2

. The launch signals Meta's aggressive push to reestablish itself as a major force in the AI race after falling behind competitors who rapidly introduced more capable models

2

.

Source: Geeky Gadgets

Source: Geeky Gadgets

Pricing and Availability Through Meta Model API and Muse Code

Muse Wonder will be accessible through the Meta Model API at $4.00 per million input tokens and $20.00 per million output tokens, making it roughly equivalent in cost to GPT-6 Sol or Opus 5.5

1

. Developers can access Wonder through the Power Plan at $20 per month or the Maximum Plan at $100 per month

3

. Muse Wonder Full, the more powerful version, will only be available to Muse Code subscribers at monthly subscriptions ranging from $50 to $200, with no public API access

3

4

. Both models will power the Muse AI agent, starting with subscribers, though Meta hasn't provided a specific timeline for wider rollout

4

.

Source: PYMNTS

Source: PYMNTS

Active Memory Feature Targets Long-Term Personal Superintelligence

Meta placed particular emphasis on Active Memory, a feature designed to help the models retain and synthesize information over extended periods. Research scientist Kavya Kopparapu stated the company "put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months and beyond"

1

. On the MemoryBench benchmark, which challenges continual learning and memory construction frameworks, Muse Wonder now scores higher than Claude Fable, Opus, or GPT-6 Astra

1

. This capability aligns with Meta's focus on developing personal assistants that can perform agentic tasks over long periods, distinguishing its approach from competitors who prioritize knowledge work and automation

5

.

Active Focus Enables Complex Task Handling and Error Recovery

The second novel feature, Active Focus, keeps the model locked into stated goals and prevents confusion during complex research or large software projects

2

. With Active Focus, Muse Wonder can pursue alternative approaches, recover from errors, and ask users for clarification when encountering technical dead ends

3

. Meta says the models deliver state-of-the-art performance on personal and professional agentic tasks across expert topics including law, finance, health, and coding

5

. The company reports Muse Wonder outperforms both Claude Fable 5.1 and Claude Opus 5.5 on specialized benchmarks measuring legal document analysis and software development capabilities

2

.

Source: Wccftech

Source: Wccftech

Meta Superintelligence Labs Delivers After Alexandr Wang Restructuring

The models represent the first major frontier release since Alexandr Wang, former Scale AI founder, restructured Meta's AI efforts in August of last year. Zuckerberg's gamble involved sinking billions into the newly formed Meta Superintelligence Labs and acquiring Scale AI, with Wang appointed as chief AI officer

5

. The effort has been rocky at times, with some researchers not remaining long-term

2

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

1

. Zuckerberg told the Financial Times achieving frontier AI models "required a large amount of compute and investment, and a lot of very rigorous scientific work by the group working on it over an extended period of time"

5

.

Safety Concerns Emerge Amid Meta's Self-Regulation Stance

Meta says it has trained Muse Wonder to refuse generating dangerous outputs such as bioweapon recipes or hacking assistance, claiming strong scores on benchmarks measuring resilience to harmful prompts

2

. Zuckerberg delayed the model release last month to work on safety and revealed a framework for scaling capabilities responsibly

2

. However, Meta did not disclose whether it has observed agents powered by Muse Wonder misbehaving in testing or specific mitigation measures, despite recent incidents involving rogue AI agents from Anthropic and OpenAI

2

. Zuckerberg has notably stayed out of the coalition of tech CEOs calling for AI development slowdowns, telling Financial Times he believes "AI is a technology that can be engineered with normal guardrails and built responsibly" rather than requiring external regulation

3

. This self-policing approach contrasts sharply with calls from Anthropic CEO Dario Amodei and OpenAI's Sam Altman for more cautious development, raising questions about whether Meta's safeguards will prove sufficient as the company pushes toward personal superintelligence for everyone

3

.

Today's Top Stories

© 2026 TheOutpost.AI All rights reserved