Meta is challenging OpenAI and Anthropic for the subscription market with the Monday launch of Muse Wonder and Muse Wonder Full, frontier-grade reasoning systems built for personal and professional agentic tasks.
Meta Deploys Muse Wonder Models
Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday. Codenamed Watermelon internally, the release is Meta’s first frontier model since Chief AI Officer Alexandr Wang and Chief Executive Officer Mark Zuckerberg led a 14-month leadership and infrastructure overhaul.
The Wonder models are larger and built for complex reasoning across coding, finance, law, and document work, whereas Meta’s earlier Muse Spark aimed to be a lightweight model operating at low latency. Meta’s model card testing shows the new software achieves benchmark scores comparable to Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra.
Alexandr Wang described the models in an interview with Bloomberg as true state-of-the-art AI systems built to handle complex personal and professional agentic work. Research scientist Kavya Kopparapu noted in an accompanying launch video that the company invested a tremendous amount of effort into ensuring the models can remember important user information over days, weeks, months, and beyond.
Active Memory and Focus Drive Muse Wonder Tasks
Meta designed the Muse Wonder family to execute multi-step workflows like booking reservations, interpreting leases, and building applications. The software relies on two core mechanisms: Active Memory and Active Focus.
Active Memory allows the assistant to retain and synthesize information across days, weeks, and months. Meta’s testing on the MemoryBench evaluation suite places regular Wonder ahead of competing models on long conversational tracking.
Active Focus is engineered to keep the assistant working through roadblocks.
“When you do real-world tasks, you’re going to get roadblocks and you’re going to get things that come up or various setbacks. We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time.”
Alexandr Wang, Meta AI chief, via CNBC
Zuckerberg noted that Meta’s strategy prioritizes personal utility through its rapidly expanding assistant ecosystem, emphasizing that the company is focused on personal empowerment rather than automation-type cases. This contrasts with rivals focused primarily on automated knowledge work.
Meta Sets Tiered Pricing for Apps and APIs
Access to the Muse Wonder models is split across consumer applications, developer tools, and specialized coding platforms.
- Meta Model API: Standard Wonder is available in preview under model ID muse-wonder-1.0, charging $4 per million input tokens and $20 per million output tokens.
- Muse Consumer App: Users of the Muse personal agent app must sign up for a monthly subscription of $20 or $100 to access advanced model capabilities.
- Muse Code: Developers and software engineers working within Muse Code can access Wonder via pay-as-you-go or standard subscription plans.
- Muse Wonder Full: The most powerful variant is restricted entirely to developers subscribing to the software engineering-focused Muse Code agent under a $50 or $200 a month plan, and is not served through the standard public API.
Wang stated that developer API pricing will remain meaningfully below leading models offered by competitors, though independent evaluations indicate that task execution costs depend heavily on workload composition.
Meta Scales Infrastructure Amid Privacy Scrutiny
Meta asserts that Muse Wonder offers strong safeguards against sharing private information without permission and refusing harmful requests.
Meta disputed claims of privacy lapses. Wang attributed the incidents to users forgetting that they authorized communications earlier in their conversations, adding that the company is educating users on such matters to ensure consumer trust.
Total capital expenditures are projected between $130 billion and $145 billion for 2026, driven by data centers, custom training chips, and contracted nuclear power capacity led by infrastructure head Santosh Janardhan. Zuckerberg maintained that every lab has the responsibility and incentive to move at the required pace for safe model training.