This week, Meta released Muse Wonder, claiming that some of its benchmark results have surpassed comparable models from OpenAI and Anthropic. Zuckerberg also revealed that the next-generation model is already in development, with future iteration speeds likely to accelerate significantly—Meta’s investment in AI shows no signs of slowing down.
Meta Platforms (META.O) launched its next-generation AI model, Muse Wonder, on Monday, stating that its performance on certain technical benchmarks has reached industry-leading levels.
Meta CEO Mark Zuckerberg said the company has begun advancing its next-generation models, with the pace of future model development "likely to be much faster," signaling continued expansion of its AI investments.
Muse Wonder will be used to upgrade Muse, the personal AI agent launched by Meta last month, and will serve as a key product for evaluating whether the company’s substantial AI investments can drive user growth and revenue.
Meta states that the new model outperformed comparable models from OpenAI and Anthropic in certain tests, with enhancements in memory, complex task handling, and finding alternative solutions after encountering obstacles.
The next round of investment after "reaching the frontier"
In an interview with the Financial Times, Zuckerberg said that Muse Wonder will support “more sophisticated behaviors at the forefront of artificial intelligence.” He noted that developing this model requires substantial computational resources and investment, as well as long-term, rigorous scientific work by the team.
Meta claims that Muse Wonder outperforms Anthropic's Claude Opus 5.5 and Fable 5.1, as well as OpenAI's GPT-6 Astra, in certain technical benchmarks.
Meta’s Chief AI Officer, Alexandr Wang, said the company believes Muse Wonder is capable of competing with today’s leading frontier models and is particularly well-suited for personal and professional agents.
The model's internal codename is "Watermelon." According to Business Insider, Alexander told employees in July that Watermelon has caught up to OpenAI, using an order of magnitude more computing resources than Meta's previous "Avocado"-codenamed Muse Spark model.
Zuckerberg said Meta "is already developing the next stage" and expects the pace of model development to accelerate significantly thereafter. This statement indicates that, after significantly increasing investments in data centers, computing infrastructure, and AI talent, Meta currently has no clear plans to reduce capital spending.
Over the past year, Meta has undergone significant organizational and resource realignments around AI. After the Llama 4 release in 2025 failed to meet company expectations, Zuckerberg restructured the AI team, established Meta Superintelligence Labs, and brought in Alexander to lead core AI research, while offering high salaries to recruit researchers from competitors and research institutions.
The new round of investments previously raised concerns among investors regarding capital expenditures. However, according to the Financial Times, since Meta launched Muse in September this year, the company’s stock has risen by nearly 10%. On the day Muse Wonder was released this Monday, Meta’s stock increased by approximately 3%.
Muse becomes a commercialization testbed
The direct application of Muse Wonder is not to provide a standalone chat model, but to enhance Muse, Meta's personal agent, in handling real-world tasks.
Alexander noted that when agents perform real-world tasks such as ordering food, checking services, or arranging travel, they frequently encounter website restrictions, missing information, or other obstacles. To address this, Meta specifically enhanced Muse Wonder’s ability to find alternative pathways during training, enabling it to persist in trying different approaches after setbacks rather than immediately stopping and requesting new instructions from the user.
Meta says the new model also enhances cross-dialogue memory and the ability to handle complex, multi-step tasks, making it suitable for understanding leases, processing financial and health information, and advanced programming. The larger Muse Wonder Full offers even greater performance.
These features will first be available in paid products. Muse users must subscribe to a $20 or $100 monthly plan to access Muse Wonder; Muse Wonder Full will be available to software developers using Muse Code, with subscription plans priced at $50 or $200 per month.
Meta also plans to offer Muse Wonder to developers via the Model API, and Alexander said its price will be "significantly lower" than the leading models from OpenAI and Anthropic.
This arrangement enables Muse Wonder to simultaneously participate in model competition and commercialization testing. Meta’s long-term revenue structure heavily relies on online advertising, and the company is exploring new AI revenue streams through subscriptions, developer APIs, and enterprise model services.
Muse has already achieved rapid initial growth. According to Sensor Tower data cited by The Wall Street Journal, the app has reached nearly 9 million global downloads since its launch. However, its download growth rate has begun to slow, and its 14-day user retention rate is lower than that of ChatGPT.
The widespread adoption of agents still faces practical limitations.
Whether the performance improvements of Muse Wonder will directly translate into user retention remains to be seen. Some limitations stem from the platform environment beyond the agent itself.
Jefferies analyst Brent Thill said he was initially impressed by Muse’s capabilities, but an increasing number of websites are blocking agents from accessing their sites, negatively impacting the user experience. At the end of September, Amazon blocked Muse from completing purchases on its site on behalf of users, and several other websites have also restricted agents from performing actions.
According to The Information, citing OpenRouter data, only 14% of AI model usage generated through the platform in the third quarter of this year came from the top ten high-quality models, down from 55% in the first quarter of last year. As lower-cost models improve in capability, companies are increasingly reserving the most advanced and expensive models for only a few high-complexity tasks.
Meta still chose to have Muse directly adopt cutting-edge models. Zuckerberg described the company’s direction as “personal superintelligence,” aiming for models to ultimately serve individuals’ personal and professional lives, rather than focusing primarily on automating knowledge work.
He also acknowledged that while artificial intelligence has rapidly improved in its ability to handle advanced academic tasks, there remains a significant gap between its current capabilities and becoming a reliable daily personal assistant.
For Meta, Muse Wonder becomes the first critical test since last year’s AI restructuring: the company has claimed a return to the forefront of model capabilities; now it must demonstrate that increased computational investment and faster model iteration can enhance Muse’s real-world user experience and drive users from downloads to sustained usage and paid subscriptions.