Meta Unveils Dual AI Models: Proactive Memory Architecture Takes on GPT-6 Astra

Meta has officially released two AI models, Muse Wonder and Muse Wonder Full, marking the company's first frontier models built on a "proactive memory" architecture — aimed squarely at the industry's long-standing challenge of long-term information retention.

According to Hacks.gr, both models were trained at the Prometheus data center, with benchmark results on par with Fable 5.1 and GPT-6 Astra. This is the first major release since Alexandr Wang reorganized Meta's AI division last August, with researcher Kavya Kopparapu presenting the details in a video.

Proactive Memory: Solving AI's "Amnesia"

One of the most glaring weaknesses of today's mainstream large models is the fragmentation of conversational memory — each interaction starts nearly from scratch, forcing users to repeatedly supply context. Meta is positioning "proactive memory" as the core selling point of Muse Wonder, emphasizing the model's ability to retain and combine user information over time.

In the company's internal memory tests, Muse Wonder outperformed systems such as Fable, Opus, and Astra. This suggests the model no longer relies solely on the passive filling of a context window, but possesses some degree of active information management capability.

This direction is not unique to Meta. The entire industry is exploring how to make AI assistants truly "remember" user preferences and historical interactions, but most approaches remain at the level of bolted-on vector databases. Whether Meta's claimed "proactive memory architecture" implies native support at the model level remains unclear given currently available public information.

Context window refers to the total amount of text a large model can "see" in a single conversation, typically measured in tokens, with anything beyond that being truncated or sliding out. Mainstream models' context windows have expanded from a few thousand tokens in the early days to hundreds of thousands or even millions today, but the fundamental nature remains "one-shot": the model does not automatically carry over content from a previous session unless it is explicitly re-written into the prompt. Bolted-on vector databases are currently the most common workaround — compressing historical interactions into vectors stored in an external database and retrieving them back into context when needed. This "retrieval-augmented" approach is flexible but passive, dependent on retrieval quality, and struggles with updating, conflicting, or expiring information. So-called native "proactive memory" generally refers to mechanisms for writing, forgetting, and prioritizing long-term information that are built into the model's architecture or training phase, rather than relying entirely on external retrieval. The difference is akin to "flipping through notes" versus "truly remembering" — and the industry has yet to reach consensus on which approach is more reliable.

Pricing Strategy and Product Tiering

Muse Wonder's API is priced at $4 per million input tokens and $20 per million output tokens, available through Meta's coding platform. This pricing sits in the mid-to-high range of the current frontier model market.

The more powerful Muse Wonder Full takes a different distribution strategy — available exclusively to Muse Code subscribers and not offered separately through the coding platform. This tiering strategy mirrors the approach of companies like OpenAI and Anthropic, binding high-value users through subscriptions while maintaining competitiveness in the API market.

Both models will power Meta's Muse digital assistant, forming a complete pipeline from underlying model to end product.

The pricing difference between input tokens and output tokens reflects the cost structure of inference. Input tokens are the prompts users send to the model; they can be processed in parallel and cached for reuse. Output tokens are generated word by word, requiring a fresh forward computation at each step, which makes compute costs higher — hence output prices are typically significantly higher than input prices. Taking this article's $4/$20 as an example, output is 5x the input, consistent with current industry norms. Developers evaluating costs should look beyond unit price and factor in the input-to-output ratio of their task: summarization tasks have large inputs and small outputs, while long-form generation is the opposite. Additionally, mechanisms like cache hits and batch calls can further drive down actual costs, and there is often a notable gap between list prices and real bills — so selection decisions should be based on stress tests with real workloads.

Data Center and Organizational Background

The Prometheus data center is located in central Ohio and began operations earlier this year. Choosing this timing for the model release reflects that Meta's continued investment in compute infrastructure is beginning to bear fruit.

Alexandr Wang's reorganization of Meta's AI division last August is an important organizational backdrop to this release. That Kavya Kopparapu appeared as a researcher to present the product suggests Meta is trying to have its technical team communicate more directly with the public, rather than relying solely on executives.

Competitive Landscape: Benchmarking Against Fable 5.1 and GPT-6 Astra

Meta is directly benchmarking Muse Wonder's results against Fable 5.1 and GPT-6 Astra — a clear competitive signal. However, there is often a gap between benchmark tests and real-world experience, especially in dimensions like memory that are difficult to quantify.

The actual effectiveness of "proactive memory" requires large-scale user validation. Whether the model can reliably retain key information over weeks or even months of interaction while avoiding privacy leaks and the accumulation of false memories — these are questions still unanswered.

Meta's release shows differentiated intent in its technical direction, but whether it can ultimately build a true moat in memory capability depends on how much architectural detail is disclosed and real-world testing feedback from developers.

Benchmarks are standardized test sets for evaluating model capabilities, with common dimensions including knowledge Q&A, mathematical reasoning, and code generation. Their value lies in being quantifiable and reproducible, but the limitations are obvious too: once questions are made public, they can enter training data, leading to inflated scores through "leaderboard gaming"; and capabilities like memory, long-term consistency, and privacy boundaries are hard to design into tests that are both objective and close to real-world use. When vendors cite benchmark results at launches, they typically choose the dimensions where they perform best — readers should note whether the test set is public and whether it was run under the same conditions as competing models. More reliable signals come from independent third-party evaluation platforms and large-scale developer testing feedback, especially how models perform across continuous multi-turn, cross-session scenarios.

Additional Background

The compute scale of a data center directly determines the training cadence of frontier models. A large AI data center typically contains tens of thousands to hundreds of thousands of accelerator cards, paired with high-speed interconnect networks and liquid cooling; a single frontier model training run may last weeks and consume tens of millions of dollars in electricity. As a result, the commissioning of a new data center often brings a wave of model releases — once compute is in place, experiments previously queued due to resource constraints can be scaled up. Site selection considerations for places like central Ohio include power supply, land costs, and local policy incentives. For observers, comparing model release timing against compute infrastructure progress provides a clearer picture of whether a given release marks a technical breakthrough or is the result of capacity coming online.

One-line Summary

Meta releases its first Muse Wonder series models with a proactive memory architecture, benchmarked against Fable 5.1 and GPT-6 Astra.

Paragraph Summary

Meta has officially released two frontier models, Muse Wonder and Muse Wonder Full — its first products built on a "proactive memory" architecture, aimed at solving the industry-wide problem of fragmented conversational memory where each interaction starts nearly from scratch. Trained at the Prometheus data center in Ohio, the models benchmark on par with Fable 5.1 and GPT-6 Astra, and outperform Fable, Opus, and Astra in internal memory tests. On pricing, Muse Wonder is available through the coding platform at $4 per million input tokens and $20 per million output tokens, while the more powerful Full version is limited to Muse Code subscribers — a classic product tiering approach. The release also marks the first major outcome since Alexandr Wang reorganized Meta's AI division last August, with both models set to power Meta's Muse digital assistant, forming a complete pipeline from underlying model to end product.