On October 12, 2026 local time, Meta officially announced on its website the new generation of flagship AI models—Muse Wonder and Muse Wonder Full. This marks Meta's first real entry into the "frontier model" category. The company stated that the two models' performance on reasoning, writing, coding and agent tasks has approached or in some cases surpassed top products from Anthropic and OpenAI, signaling a significant narrowing of the gap with industry leaders in the AI race.

Details

The Muse Wonder series was led by Meta Superintelligence Labs and trained at the company’s gigawatt-scale Prometheus data center. The models continue the pretraining and reinforcement learning trajectory of the earlier Muse Spark and have been scaled up substantially. Both models support a 1,000,000-token context window and can process mixed inputs of text, images, video and audio.

Key capability highlights include two agent-oriented mechanisms:

Active Memory: can retain and recall user-relevant information across conversations, maintaining contextual coherence for long-running tasks.
Active Focus: when a task is blocked, the model attempts multiple paths, recovers from errors, proactively clarifies and confirms before critical steps, reducing the risk of agents "drifting" off-task.

Meta emphasized these models are designed for personal and professional agent use cases across complex domains such as law, finance, healthcare and programming—for example, handling insurance claims, interpreting leases, and building runnable applications from scratch.

Benchmark performance (according to Meta's model card and independent evaluator Artificial Analysis):

On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full scored 56 (highest reasoning intensity), matching Claude Sonnet 5.5 and trailing only Claude Opus 5.5 (58), while outperforming GPT-6 Astra (53) and Claude Fable 5.1 (53).
Muse Wonder scored 52, a notable improvement over the previous-gen Muse Spark 1.3 which scored 48.
Strong performance on long-horizon coding tasks: on DeepSWE 1.1 Wonder Full reached 79.0%, and on SWE-Marathon 57.5% (versus Claude Opus 5.5 at 46.3%).
Leading or top-three results on multiple agent-focused evaluations such as the legal agent benchmark Harvey LAB-AA and the automation workflow AutomationBench-AA.
On memory-related benchmarks (MemoryBench), the Wonder series also exceeded major competitors.

It should be noted that some coding comparisons were not fully controlled (different models used different scaffolding), and knowledge-based evaluations remain a comparatively weaker area.

Availability and pricing：

Muse Wonder: available now via the Meta Model API and Muse Code. API pricing is $4 per million input tokens, $20 per million output tokens, and cached input at $0.20, roughly comparable to GPT-6 Sol or Claude Opus 5.5.
Muse Wonder Full: currently available only to Muse Code subscribers (monthly fee around $50–200) and not offered via a public API. Meta plans for Wonder to gradually power personal agent products (Muse), prioritizing subscribers.

Neither model's weights are available for download, continuing Meta's recent shift from open-source Llama releases toward closed-source commercialization.

Potential impact

1. Meta re-enters frontier competition

Previously, Meta held an important position in the open-source ecosystem with the Llama series—especially Llama 3—but after Llama 4 underperformed expectations the company carried out large AI hiring and reorganizations. Despite that, Meta long trailed OpenAI and Anthropic on closed-source flagship performance. The Muse Wonder series launch,