Meta Launches Muse Wonder Series of Cutting-Edge Models, Claims They Are the Company's Most Powerful Models to Date
Meta launched the Muse Wonder and Muse Wonder Full series of models on October 12, claiming they are the company's most powerful models to date, achieving industry-leading levels in personal and professional agent tasks.
The new models enhance capabilities in specialized fields such as law, finance, and healthcare, as well as advanced programming capabilities. They introduce Active Memory functionality to retain key information across conversations, and Active Focus functionality to continuously advance in complex tasks, explore different paths, correct errors, and follow up on execution. The models can handle multi-step tasks ranging from insurance claim navigation, understanding lease documents, to building runnable applications, and support text, image, video, and audio inputs, with a context window of 1 million tokens.
Muse Wonder is available starting today through the Meta Model API and Muse Code, priced at $4 per million input tokens, $0.2 for cached input, and $20 for output; Muse Wonder Full is only available to Muse Code subscribers and is not sold through a public API. More Muse products will be supported in the future, initially targeting subscribers to drive personal agent experiences.
This series builds on Muse Spark and is the latest achievement since the company initiated an AI business restructuring 14 months ago, developed by Superintelligence Labs, with training completed at the Giga-level Prometheus data center. Unlike the previously open-source Llama series, the Wonder series does not release downloadable weights and uses a closed API and subscription model.
Independent benchmark testing shows that Muse Wonder Full scored 56 on the Artificial Analysis Intelligence Index, ranking alongside some leading models and surpassing some competitors, with significant improvements in agent and coding-related evaluations; Muse Wonder scored 52, up from 48 for the previous Muse Spark 1.3. The company's stock price rose about 4% following the announcement.
Market Mechanism: Meta charges developers and professional users for cutting-edge model capabilities through APIs and subscription tools, driven by the demand to catch up with competitors like OpenAI and Anthropic following previous AI restructuring and computing power investments; funds flow from enterprises and subscription users to Meta's model services and subscription products, benefiting Meta (new API and subscription revenue streams) and developers using professional agents, while competitors relying on open-source or low-cost models face pressure (facing competition from closed high-performance model pricing).
Source: Public Information
ABAB AI Insight
Previously, Meta promoted the proliferation of downloadable models through the open-source Llama series, but around 2025, it acquired Scale AI for approximately $14.3 billion and brought in Alexandr Wang to lead Superintelligence Labs, initiating a 14-month AI restructuring that shifted from foundational rebuilding to Giga-level data center training and cutting-edge agent capabilities.
In terms of capital pathways, the company is converting large-scale computing power and talent investments into closed API pricing ($4 per million input tokens, $20 for output) and tiered Muse Code subscriptions, limiting the strongest model Wonder Full to paying users, aiming to directly convert agent task capabilities into measurable enterprise and professional revenue, rather than continuing to distribute value in open-source form.
This is analogous to OpenAI's shift from research models to ChatGPT subscriptions and API charges, or Anthropic locking in enterprise and coding scenarios with the Claude series; currently, Meta is transitioning from an open-source ecosystem participant to the commercialization of closed cutting-edge models, attempting to establish pricing power alongside OpenAI and Anthropic in the personal and professional agent space.
Essentially, this represents a technological substitution under the transfer of pricing power: agent and long-range task capabilities become new sources of differentiation, forcing models to shift from downloadable open-source assets to closed services charged by token and subscription, with the mechanism being to recover high computing training costs through APIs and subscriptions, while locking in user workflows with features like Active Memory and Active Focus, reducing the likelihood of migration.
ABAB News · Cognitive Laws
Open-source proliferation models, closed pricing capabilities.
Agent tasks turn computing costs into measurable subscription revenue.
The strongest models are first reserved for paying users, then a decision will be made on whether to open them.