AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Meta releases its first frontier model, Muse Wonder

Collected Oct 12, 2026

Meta's Superintelligence Lab released a pair of state-of-the-art models on Monday: Muse Wonder and Muse Wonder Full. The company describes them as its first attempt at a frontier model, a category defined by the largest, most capable systems labs build at the leading edge of capability.

The two models differ in scope. Wonder is served through the Meta Model API, while Wonder Full is reserved for Muse Code subscribers and will not be offered through the API at all. Both will also be used to power the Muse agent, Meta's assistant product.

The release follows Muse Spark, which Meta positioned as a lightweight model built for low latency. By contrast, the Wonder models are larger and are described as more capable, with higher benchmark scores and better results on complex tasks such as coding. Meta has been shipping smaller pieces of its AI stack for a while, so the Wonder launch marks a shift toward competing directly on scale rather than on efficiency alone.

According to Meta's own testing, presented in the model card, Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra. Meta also highlighted the MemoryBench benchmark, which it says challenges continual learning and memory construction frameworks. On that benchmark, Muse Wonder scores higher than Fable, Opus, or Astra, per the company's testing.

Pricing reflects the larger footprint. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens, which the company frames as roughly equivalent in cost to GPT-6 Sol or Opus 5.5. The absence of an API path for Wonder Full means most developers will interact with the smaller of the two models, with the full variant effectively bundled into a subscription product rather than sold per token.

A central theme of the launch is memory. Meta emphasized the models' ability to retain and synthesize information supplied by the user, a property it calls active memory. That emphasis lines up with the company's stated interest in personal assistants that can carry out tasks over long periods. Research scientist Kavya Kopparapu said in a video accompanying the launch that a tremendous amount of effort went into making sure the model can remember what is important to a user over days, weeks, months, and beyond.

The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year, according to the company's post.

The launch is also notable for timing. It is the first frontier model release since Scale AI founder Alexandr Wang restructured Meta's AI efforts in August of last year. Since that restructuring, Meta shipped the lightweight Muse Spark model and the Muse agent but had avoided the kind of state-of-the-art releases OpenAI and Anthropic are known for. Meta's blog post describes the Wonder pair as the next step on its scaling ladder and a major leap toward personal superintelligence.

In practice, the benchmark claims matter less than the packaging. Comparable scores to Fable 5.1 and GPT-6 Astra, on Meta's own testing, put the Wonder models in the same conversational tier as the systems developers already evaluate. Self-reported results are the norm at launch, and independent evaluation typically takes weeks to arrive, so treating parity as provisional is the sensible reading.

The cost structure is the more concrete comparison. At $4 per million input tokens and $20 per million output tokens, Wonder sits near GPT-6 Sol and Opus 5.5, so Meta is not undercutting rivals on price. What it is offering instead is a memory-first positioning: long-horizon recall aimed at assistants that act across sessions rather than single-turn chat.

That framing carries a trade-off. Long-horizon memory is useful for agents that track ongoing projects, but it also raises questions about how user data is stored, how recall is scoped, and how stale or incorrect memories get corrected. Meta has not detailed those mechanics in the launch material.

Why it matters: Teams choosing a frontier model now have another option at roughly the same per-token cost as incumbent systems, and the differentiator Meta is pushing is memory rather than raw benchmark leadership. The split between Wonder and Wonder Full means the API-accessible model is the one most product teams can plan around, while the strongest variant is gated behind a subscription and cannot be wired into external applications. Developers building long-running agents should watch independent evaluations and the details of how active memory is managed before committing.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Where Muse Spark aimed to be a lightweight model that could perform at low latency, the Wonder models are larger and said to be more capable, with higher benchmark scores and better results on complex tasks like coding.