Meta has launched two new artificial intelligence models designed to compete head-on with OpenAI and Anthropic, in a move that marks a dramatic shift away from the open approach that once defined its AI strategy.
The company said Muse Wonder and Muse Wonder Full are its most capable models to date. They are built to power personal and professional AI agents that can handle tasks from software development to document analysis and financial research.
Fourteen months of rebuilding
The launch follows about 14 months of work to rebuild Meta's AI capabilities under Meta Superintelligence Labs, led by chief AI officer Alexandr Wang. The models were developed using Meta's Prometheus data centre and techniques spanning pre-training and reinforcement learning.
Meta has spent tens of billions of dollars on AI infrastructure and talent over the past two years, after its earlier models were seen to fall behind rivals.
Closed doors
The release also marks a change of philosophy. Meta's earlier Llama models were "open-weight", meaning developers could download them, run them on their own computers and adapt them freely. That made Meta a champion of open AI.
Neither Muse Wonder model is available for public download or self-hosting. Access is instead provided through Meta's own services and platforms.
Supporters of open models have warned that the shift reduces competition and transparency. Others argue that the most powerful systems carry risks that make tighter control sensible.
Memory and focus
Meta says the new models include two features intended to make AI agents more useful over long tasks.
"Active Memory" lets the models retain relevant information across conversations and retrieve it when needed. "Active Focus" is designed to help them work through obstacles, try alternative approaches, recover from errors and ask for clarification when necessary. The models are also designed to seek confirmation before taking important actions.
These features target a well-known weakness of AI agents. Completing a multi-step task requires more than producing a convincing answer: an agent must track requirements, use tools correctly and adapt when its original plan fails.
Meta also claims advanced coding performance and specialist knowledge in areas including law, finance and health.
How it compares
Independent evaluations by Artificial Analysis give Muse Wonder Full a score of 56 on its Intelligence Index, tying it with Anthropic's Claude Sonnet 5.5. Muse Wonder scores 52, while OpenAI's GPT-6 Astra scores 53.
The results suggest Meta has become far more competitive with leading AI developers. But the models do not lead every benchmark, and analysts caution that test scores do not guarantee reliable performance in real business settings.
Privacy questions
The memory features are likely to raise privacy concerns. AI systems that remember information across conversations could store sensitive personal or business data, and regulators in Europe have already scrutinised Meta's use of user data to train AI.
Businesses considering the models will want clear answers on where data is stored, how long it is kept and whether it is used to improve future systems.
A crowded race
The launch comes during an intense week in the AI industry. Google has made its latest Gemini model widely available, and debate is growing over safety, after Microsoft chief executive Satya Nadella called for an "emergency brake" on advanced systems.
Pricing for business customers has not been fully detailed, and Meta has said access will initially be limited while it expands capacity, a sign of how much computing power the new models require.
For Meta, which reaches billions of people through Facebook, Instagram and WhatsApp, the stakes are high. If Muse Wonder delivers, it could put powerful AI agents in front of more users than any rival. If it falls short, the company's enormous investment will face hard questions from investors.
