Meta launches Muse Wonder, its first frontier AI model with active memory

Meta’s Superintelligence Lab has officially entered the frontier model arena with the release of Muse Wonder and Muse Wonder Full. Announced on Monday, October 12, 2026, these models represent the company’s first attempt to compete directly with state-of-the-art systems from other major AI labs. The launch signals a strategic shift for Meta, moving beyond lightweight efficiency models toward high-capability systems designed for complex reasoning and long-term interaction.

What happened

The new releases, Muse Wonder and Muse Wonder Full, are positioned as larger and more capable than Meta’s previous Muse Spark model. While Muse Spark focused on low latency and lightweight performance, the Wonder variants target higher benchmark scores and superior results in demanding tasks such as software coding. According to Meta’s internal testing documented in the model card, both Wonder models achieve performance metrics comparable to Fable 5.1 and GPT-6 Astra. This places them firmly in the top tier of currently available commercial AI models.

Access to these models is structured through different channels. Muse Wonder is available via the Meta Model API, priced at $4 per million input tokens and $20 per million output tokens. This pricing strategy aligns it closely with competitors like GPT-6 Sol and Opus 5.5. In contrast, Muse Wonder Full will not be served through the public API. Instead, it is exclusively available to subscribers of Muse Code, Meta’s specialized coding environment. Both models will also serve as the underlying engine for the Muse agent, Meta’s broader personal assistant platform.

This launch follows a significant restructuring of Meta’s AI efforts led by Scale AI founder Alexandr Wang in August of the previous year. Since that reorganization, Meta had focused on releasing lighter models and agents but had avoided the high-profile, state-of-the-art releases typical of OpenAI and Anthropic. The training for these new models took place at Meta’s Prometheus data center in central Ohio, which became operational earlier in 2026.

How it works

A central technical innovation in the Muse Wonder models is a feature Meta calls “active memory.” This capability allows the models to retain and synthesize information provided by users over extended periods, ranging from days to months. This design addresses a common limitation in current large language models, which often struggle to maintain context or recall specific user preferences across long-running sessions. By improving continual learning and memory construction, the models aim to function more like personalized assistants rather than transient query processors.

To validate this capability, Meta utilized the MemoryBench benchmark, which specifically tests frameworks for continual learning and memory retention. In these tests, Muse Wonder reportedly scored higher than Fable, Opus, and Astra. The architecture prioritizes the ability to distinguish and store important user-specific data, enabling the model to build a persistent understanding of individual needs. As research scientist Kavya Kopparapu stated, the team focused heavily on ensuring the model can remember what matters to the user over time.

Key details

• Muse Wonder and Muse Wonder Full are Meta’s first frontier models, surpassing the capabilities of the earlier Muse Spark.
• Benchmark scores for both models are comparable to Fable 5.1 and GPT-6 Astra, according to Meta’s model card.
• Muse Wonder is available via API at $4 per million input tokens and $20 per million output tokens.
• Muse Wonder Full is exclusive to Muse Code subscribers and is not accessible through the public API.
• The models feature “active memory,” scoring higher than competitors on the MemoryBench benchmark for long-term recall.
• Training was conducted at the Prometheus data center in central Ohio, which opened earlier in 2026.

Why it matters

For software engineers and technical founders, the introduction of Muse Wonder expands the competitive landscape for high-end AI inference. The pricing parity with established leaders like GPT-6 Sol suggests that cost will no longer be the primary differentiator among top-tier models. Instead, differentiation will likely come from specialized features such as active memory and integration depth. Developers building applications that require long-term user context, such as personalized coaching apps or complex project management tools, now have a viable alternative that explicitly optimizes for memory retention.

The restriction of Muse Wonder Full to Muse Code subscribers also highlights a growing trend of vertical integration in AI services. By keeping its most powerful model within a specific subscription ecosystem, Meta is incentivizing developers to adopt its full suite of tools rather than just using raw API access. This move could influence how teams choose their development environments, balancing the flexibility of open API access against the potential performance gains of integrated, proprietary stacks.

What you can do

• Evaluate Muse Wonder via the Meta Model API if your application requires high-performance inference with competitive pricing.
• Test the “active memory” capabilities using MemoryBench-style prompts to see if long-term context retention improves your user experience.
• Consider subscribing to Muse Code if your team relies heavily on AI-assisted coding and wants access to the more powerful Muse Wonder Full.
• Compare the total cost of ownership against GPT-6 Sol and Opus 5.5, factoring in both token prices and integration complexity.
• Monitor the performance of the Muse agent powered by these models to understand how active memory translates to real-world assistant tasks.
• Review the model card for detailed benchmark data to ensure the models meet your specific accuracy requirements for coding and reasoning tasks.