Meta's Muse Wonder Is Its First Frontier Model, and It Is Closed and Priced Like the Leaders
Six months after Muse Spark, Meta Superintelligence Labs ships bigger models built around long-term memory, sold at $4 and $20 per million tokens.
Key takeaways
- Muse Wonder and Muse Wonder Full are Meta's first frontier-class models, trained at its Prometheus data center in Ohio.
- Meta prices Wonder at $4 input and $20 output per million tokens; Wonder Full is limited to Muse Code subscribers.
- Meta says the models match rival flagships on its own tests; no specific scores were published in launch coverage.
On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, which Meta calls its first frontier models. They are larger and more capable than April's Muse Spark, centered on a feature Meta calls active memory, and, like Spark, they are sold through Meta's own API and subscriptions rather than released as open weights.
What Meta shipped
Muse Spark, launched April 8, was the first model from the lab led by Alexandr Wang and a deliberate departure from Llama: natively multimodal, optimised for low latency, and the first Meta flagship without public weights. Wonder is the heavier follow-up. Meta says it performs better on complex work such as coding, and that Wonder and Wonder Full score comparably to rival flagships on the company's internal tests. TechCrunch reported no specific numbers, so outside readers have little to verify yet. The one benchmark Meta singled out was MemoryBench, a memory and continual-learning test, where it says Wonder beats rivals' top models.
Active memory is the pitch. The models retain and synthesise what users tell them over long periods, which Meta describes as a major step toward personal superintelligence. Both models power the Muse agent. Wonder Full is restricted to Muse Code subscribers and is not offered through the API, while Wonder is available to developers through the Meta Model API. Meta trained them at the Prometheus data center in central Ohio, which came online earlier this year.
Priced like a flagship
Wonder costs $4 per million input tokens and $20 per million output tokens through Meta's API, which TechCrunch described as roughly in line with the top models from OpenAI and Anthropic. A September roundup of frontier launches shows the neighbourhood: Anthropic's Sonnet 5.5 and Google's Gemini 4 Argon at an introductory $2 and $10, with Argon scheduled to rise to $4 and $20, and Mistral's gated Large 4 preview at $1.36 and $4.18. Meta is pricing at the upper end rather than undercutting, a statement that it sees its models as peers.
Why it matters
Meta spent 2023 to 2025 as the champion of open weights, with Llama setting the pace for the whole ecosystem. That identity faded after Llama 4 disappointed and the company reorganised under Wang. Muse Wonder completes the turn: Meta now competes as a conventional frontier lab, selling tokens and subscriptions, and has said only that it hopes to open future versions. For the open-weight race, the leadership has shifted to Chinese labs such as DeepSeek, Alibaba and Moonshot.
The memory emphasis is also strategic. Meta owns distribution through its apps and a trove of social context, and a model that remembers users across conversations is a product advantage that a pure API vendor cannot easily replicate, though it brings obvious privacy questions about what is retained and for how long.
The competitive landscape Meta is entering
The release lands in a crowded autumn. By one roundup, OpenAI shipped GPT-6 Astra on September 3 and followed with GPT-6 Sol and Luna later in the month, Anthropic released Claude Opus 5.5 and Sonnet 5.5 on September 22 and 28, and Google put Gemini 4 Argon out on September 30. Mistral opened a gated preview of Large 4, a 1.05-trillion-parameter model with 49 billion active parameters, on October 6. Meta is therefore the latecomer among the large Western labs, and its claim of parity rests on its own testing.
That matters because Meta lacks the enterprise sales channel of Microsoft or Google Cloud, and its API is new. Its advantages are scale of capital, owned data-centre capacity and billions of daily users. The Muse agent and Muse Code subscription give it paths to monetise that do not depend on winning API share.
What to watch
- Independent scores for Wonder on coding and agent benchmarks, where Spark trailed rivals.
- How Meta handles memory controls, retention and data deletion for active memory.
- Whether the promised open-weight versions ever appear.
- Developer uptake of the Meta Model API against established providers at similar prices.
Sources
- Meta releases its first frontier model, Muse Wonder — TechCrunch
- Meta unveils new frontier-class AI models to compete with OpenAI, Anthropic — Axios
- Meta's Muse Spark is its first frontier model and its first without open weights — The Decoder
- Google, OpenAI, Anthropic and Mistral release new frontier models in late September — DutchStartup.ai
