Meta Launches Muse Wonder: Paid API, Muse Agent Limited to the U.S.

Meta makes Muse Wonder accessible via its Meta Model API and in Muse Code, with a pay-per-use pricing model. The most powerful version remains exclusive to subscribers, while the Muse agent is still unavailable in Europe and has no announced timeline. The new models bring agent capabilities, a context window of one million tokens, and enhanced safeguards.

Availability, Pricing, and Access Scope

Muse Wonder is accessible via the Meta Model API and integrated into Muse Code. The pricing is set at $4 per million tokens for input and $20 per million for output, whereas Muse Spark 1.2 was offered at $1.25 and $4.25 at the launch of Muse Code. Muse Wonder Full, the most powerful version, is not available via the API and remains reserved for Muse Code subscribers. Meta anticipates that Muse Wonder will soon power the Muse agent, initially for subscribers, with no specified date. The company also plans to use Muse Wonder for artifact generation and the development of specialized models for Muse users. To date, the Muse agent has been available in the United States since September, but it is not offered in France or the European Union, and no deployment date has been communicated. For developers, the announcement does not mention any geographical restrictions regarding the Meta Model API and Muse Code.

Agent Capabilities and Supported Formats

Muse Wonder and Muse Wonder Full are designed for agent tasks, whether personal or professional. Meta highlights fields such as law, finance, and healthcare, with usage examples ranging from reading a lease to dealing with an insurer. The company claims a leading edge in programming. The models can handle various types of inputs, including text, images, videos, and audio, and feature a context window of one million tokens. According to Meta's research blog, the base version, which is the largest, outperforms others on tasks that were not covered during training. In demonstrations, Muse Wonder built a 3D cooking game from images and organized a wedding seating plan according to constraints.

Active Memory, Focus, and Safeguards

Active Memory allows for retaining important information about the user from one conversation to another and reusing it at the right moment. Active Focus enables the model to maintain pursuit of a goal even in the face of obstacles, adapt its strategy in case of failure, correct its own mistakes, and seek validation before any major operation. Meta states that it has trained its models not to disclose personal information without consent and to resist malicious instructions. According to a ranking by Surfshark, Muse is among the AI applications that collect the most personal data. Meta also presents its most advanced protective measures to date against diversions related to chemical, biological, radiological, or nuclear weapons, as well as cyberattacks.

Market Comparisons and Internal Trajectory

Meta now compares its models to OpenAI's GPT-6 Astra as well as Anthropic's Claude Fable 5.1 and Claude Opus 5.5. Until now, the group primarily used Muse Spark, a lighter model that powers the Muse agent. Meta presents this new generation as a major leap towards personal superintelligence. This milestone comes just over a year after the launch of its AI business overhaul.