Meta has entered the frontier AI race in earnest. On October 12, 2026, the company's Superintelligence Labs unveiled Muse Wonder and Muse Wonder Full, described as its most capable AI models to date and its first genuine attempt at a frontier-class system — the tier of models that includes OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. The announcement, made by Meta's AI chief Alexandr Wang, has been trending per news coverage across technology media worldwide and marks a strategic pivot for a company that spent years insisting its AI future would be open and free.

What exactly was announced?

The Muse Wonder family has two members. The standard Muse Wonder (API identifier muse-wonder-1.0) is available through the Meta Model API and OpenRouter as well as to Muse Code subscribers. Muse Wonder Full is the larger, maximum-reasoning variant built for the hardest, longest-running jobs — and it is not on the API at all; it is reserved for subscribers of Muse Code's $50-per-month "power usage" or $200-per-month "maximum usage" tiers. Both models will also power the Muse agent going forward, the personal-assistant app Meta released earlier this year.

The release follows Muse Spark 1.3, which arrived about a month earlier and was designed to be small, fast and low-latency with improved coding and agentic performance. Wonder goes the opposite direction: bigger and more expensive to run, but scoring higher on benchmarks and handling the complex coding and multi-step agent tasks that Spark was never built for. According to Meta, the models were trained at a gigawatt-scale data center the company calls Prometheus, using advances in both pretraining and reinforcement learning that build on the approach introduced with Spark.

The headline capabilities: Active Memory and Active Focus

Meta has placed a major emphasis on one theme: memory. The new models introduce two capabilities the company calls Active Memory and Active Focus. Active Memory retains a user's conversation history, preferences and habits across sessions, while Active Focus carries complex or ambiguous tasks through to completion without losing track of the goal.

To demonstrate the memory work, Meta highlighted the MemoryBench benchmark, which tests continual learning and memory-construction frameworks. According to the company's own testing, Muse Wonder now scores higher on MemoryBench than Claude Fable 5.1, Claude Opus 5.5 or GPT-6 Astra. "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond," research scientist Kavya Kopparapu said in a video accompanying the launch. Context window is reported at roughly one million tokens, though Meta has not officially confirmed that figure.

How does it benchmark against OpenAI and Anthropic?

Meta's own model card places Wonder and Wonder Full in the same range as the current frontier leaders — but independent evaluation paints a nuanced picture:

  • Artificial Analysis Intelligence Index: Muse Wonder scored 52, Muse Wonder Full scored 56 — placing Full just behind the current leader, Claude Opus 5.5, at 58, and tying for second place overall.
  • SWE-Marathon (long-running software engineering tasks): Muse Wonder Full scored 57.5%, ahead of Claude Opus 5.5 at 46.3% and GPT-6 Astra at 42.5%.
  • DeepSWE v1.1: Muse Wonder Full achieved 79%, ahead of GPT-6 Astra's 74.1% and Claude Opus 5.5's 74.2%.

Meta said its models now match the top models from OpenAI and Anthropic on some capabilities — but the claims have not gone unchallenged. Bindu Reddy, CEO of Abacus.AI, which runs its own model evaluations, publicly argued that Muse Wonder was "nowhere near Fable or Astra level" and said detailed benchmark results would follow, a reminder that vendor-reported scores always deserve independent verification.

Pricing: the end of the free-and-open era

Perhaps the most consequential part of the announcement is commercial. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens — landing close to what OpenAI and Anthropic charge for their flagship models, and a world away from the free, open-weight Llama era. The move adds a paid monetization push to a strategy that had previously centered on free, open AI through the Muse service. Subscribers to the Muse agent get priority access to Wonder, with free and paid tiers differentiated by model performance as well as usage limits.

It is a decisive break: for two years, Meta insisted it would stay open-source while Google and OpenAI charged for frontier access. Muse Wonder's pricing and proprietary license tell a different story, aligning Meta with the closed frontier-model business model of its rivals.

Why this matters

After Llama's poor showing against the frontier, Meta spent billions of dollars on AI infrastructure and talent over the past year, yet its models largely remained behind those of OpenAI, Anthropic and Google. The Wonder family is the company's first evidence that the gap may finally be closing — and investors noticed: Meta shares closed the regular session up 3.38% at $742.96 on the announcement. In an interview with the Financial Times timed to the launch, CEO Mark Zuckerberg said the company's AI capabilities had caught up with rivals and that Meta would "significantly accelerate the pace of model development," adding that the company is already working on the next stage.

The launch also intensifies the race toward what Meta calls "personal superintelligence" — AI agents that act autonomously across apps over long time horizons. It arrives in the same week that Anthropic warned, separately trending per news coverage, that AI models could soon "replicate" across home devices — a sign of how fast agent capabilities are colliding with safety concerns. For readers following that story, see our explainer on Anthropic's AI worm warning. The Muse agent slowdown in downloads that preceded this launch also gives Meta a reason to make the underlying model dramatically better: a smarter Wonder is the engine the whole assistant strategy depends on.

What happens next?

Expect three things in the coming weeks: independent benchmark suites publishing their own Wonder results (which will settle the debate over Reddy's skepticism), developers testing the API against Claude and GPT-6 pricing for coding workloads, and the Muse agent quietly getting smarter as Wonder rolls out behind it. The bigger question is structural: whether Meta, having joined the closed frontier-model club, can sustain the pace Zuckerberg promised — and whether users trust an AI that remembers them for months. Both will be answered well before the next model announcement.

Sources: Reuters/Yonhap via Seoul Economic Daily; TechCrunch; Cointelegraph; Neowin; AI Daily Post; Dow Jones Newswires via Morningstar (reporting on the October 12, 2026 launch and subsequent independent coverage).