Meta Releases Muse Wonder, Its First Frontier Models, With API Pricing at $4 per Million Tokens

Meta released Muse Wonder and Muse Wonder Full on October 12, the first AI models it describes as frontier-class. Meta says they are competitive with the best from OpenAI and Anthropic, but the evidence so far is Meta's own benchmark testing and a model card, with no independent results yet (MLQ, DevX).

What Meta released

The two models come from Meta Superintelligence Labs and sit above Muse Spark, Meta's smaller, lower-latency model. Meta calls them "the next step on our scaling ladder," and Axios reports they were code-named Watermelon. They were trained at Meta's Prometheus data centre in central Ohio, which Axios describes as gigawatt scale. Both will power Meta's Muse agent, and Muse Wonder will drive future Muse assistant experiences, initially for paid subscribers.

Pricing and access

Model	Access	Price
Muse Wonder (muse-wonder-1.0, preview)	Meta Model API, Muse Code	$4 per million input tokens, $0.20 cached input, $20 output; 1M-token context
Muse Wonder Full	Muse Code subscriptions only, no API	Included in $50 and $200 a month Muse Code plans

Pricing details are from launch coverage, including Yotta Labs. TechCrunch says the API price is roughly in line with OpenAI's GPT-6 Sol and Anthropic's Opus 5.5, which suggests Meta wants developers to treat Muse Wonder as a direct alternative rather than a cheaper option.

The performance claims

Meta says the models reach state-of-the-art results on personal and professional agent tasks, memory and coding, and perform well on legal, finance and health work. Per Meta's model card, Wonder and Wonder Full score comparably with Fable 5.1 and GPT-6 Astra, and on MemoryBench, which tests memory and continual learning, Meta says Muse Wonder beats Fable, Opus and Astra. Secondary coverage reports that the base Wonder trails Opus 5.5 on the Artificial Analysis Index (52 to 58), while Wonder Full reportedly beats Opus 5.5 on six benchmarks, mostly long-horizon coding, legal and finance tasks. These comparisons rest on Meta's own tables.

Alexandr Wang, who heads Meta Superintelligence Labs, told Axios the models are "competitive with the best models in the world." That is a company statement, not a measured result.

Memory and agents

Meta's main pitch is Active Memory, which retains important information a user provides over time and retrieves it when relevant, and Active Focus, which keeps an agent working toward a goal and tries different approaches when it hits an obstacle. Both target long-running agent work rather than single chats.

Safety and open source

Meta says it strengthened safeguards against cyber and biological misuse and added protections for personal information, with a fuller report to follow the model card. Meta has not said whether any Wonder model will get open weights. It has released open-weight models separately, including Muse Spark Open days later, and Wang said Meta is "still committed to open source."

For developers in India, the API is billed in US dollars, and at $4 and $20 per million tokens Muse Wonder costs far more than open or budget models such as DeepSeek's Flash, so it makes sense mainly where its agent and memory features matter.

What to watch next

Independent evaluations on coding and agent tasks, the API leaving preview, Meta's fuller safety report, and the date the Muse assistant starts using the model.

Related on Gadgets365

• Meta's Muse Spark Open is an 816B open-weight model built to run on two Mac Studios
• Google unveils Gemini 4 Argon and claims a benchmark lead
• FTC opens first US probe into AI agent risks