Meta Launches Muse Wonder Models With Coding Upgrades and “Active Memory”

TL;DR

• Meta released Muse Wonder and Muse Wonder Full on Monday, marking its Superintelligence Lab’s first frontier-model launch.

• The larger models target complex tasks such as coding, expanding beyond the lightweight Muse Spark.

• Meta reports benchmark results comparable to Fable 5.1 and GPT-6 Astra.

Meta’s Superintelligence Lab has released Muse Wonder and Muse Wonder Full, two larger AI models designed to improve coding, complex task execution and long-term recall.

The Monday launch marks the lab’s first push into frontier models following its earlier release of Muse Spark, which prioritized lightweight operation and low latency.

Meta positions the Wonder models as a step toward personal AI assistants that can retain useful information and carry out work over extended periods.

Wonder Models Target More Demanding Tasks

Muse Wonder and Wonder Full are larger and, according to Meta, more capable than Muse Spark.

The company’s model card reports benchmark scores comparable to Fable 5.1 and GPT-6 Astra, alongside stronger performance on complex tasks such as coding.

Those comparisons come from Meta’s own testing. They indicate how the company positions its models against competitors, rather than establishing an independently verified lead across all applications.

The release broadens Meta’s lineup from a model focused on responsiveness to options intended for more demanding workloads.

API Pricing Starts at $4 Per Million Input Tokens

Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens.

The pricing places it roughly alongside GPT-6 Sol and Anthropic’s Opus 5.5, according to the launch report.

Wonder Full has a more restricted distribution model: it will be available only to Muse Code subscribers and will not be offered through the API.

Both models will also power Meta’s Muse agent, connecting the new capabilities with its consumer-facing assistant.

“Active Memory” Focuses on Long-Term Recall

Meta highlighted the models’ ability to retain and synthesize information provided by users, describing the capability as “active memory.”

The company said Muse Wonder outperformed Fable, Opus and Astra in its testing on MemoryBench, a benchmark examining continual learning and memory construction.

Research scientist Kavya Kopparapu said the team focused on helping the model remember information important to users over days, weeks and months.

That emphasis reflects Meta’s ambition to build assistants that remain useful across ongoing tasks and conversations, rather than treating every interaction as isolated.

Launch Follows Meta’s AI Restructuring

The Wonder models are the first frontier-model release since Alexandr Wang restructured Meta’s AI efforts in August 2025.

Since that reorganization, Meta has introduced Muse Spark and the Muse agent while avoiding the larger frontier-model launches associated with rivals such as OpenAI and Anthropic.

Meta said the new models were trained at its Prometheus data center in central Ohio, which began operating earlier this year.

The launch gives Meta a more capable model lineup as it pursues assistants built around coding, task execution and persistent user context.