On October 12, Meta announced the release of Muse Wonder and Muse Wonder Full, two new artificial intelligence models positioned to directly challenge OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. The announcement marks a decisive moment in the frontier AI race: after months in which industry attention appeared divided among the big three, Meta has finally unveiled a model whose benchmark numbers compete shoulder to shoulder with the market leaders.
The Muse Wonder Full was trained at Meta's Prometheus data center, a facility with gigawatt-scale capacity designed specifically for training large language models. According to the company, the model has achieved state-of-the-art results in personal and professional agent evaluations—the core niche of Meta's current strategy, which is betting on AI assistants as the next major vehicle for mass adoption.
Meta's approach distinguishes itself from competitors through two key concepts it coined to describe the model's internal operation: Active Memory and Active Focus. Active Memory is designed to maintain relevant context across conversations, automatically retrieving information important to the user when relevant to the current task. Active Focus describes the model's ability to sustain effort toward complex goals—trying alternative approaches, working through blockers, and adapting to new information during extended tasks. These two capabilities are evaluated through FocusBench and SWE-Marathon, specific benchmarks for long-horizon agentic tasks.
In Meta's published benchmarks, Muse Wonder Full achieved a score of 56 on the Artificial Analysis Intelligence Index, just two points behind Claude Opus 5.5, which holds the #1 spot at 58. On SWE-Marathon, Meta's model reached 57.5% against 46.3% for Opus 5.5 and 42.5% for GPT-6 Astra. On DeepSWE 1.1, the reported figures were approximately 79% for Muse Wonder Full versus 74.2% for Opus 5.5. In contrast, Claude leads on ProgramBench (70.5 vs 50.0) and Terminal-Bench (59.6 vs 52.0), suggesting that the advantage depends on the type of evaluation.
A revealing aspect of this contest is the question of access. Muse Wonder Full—the variant that achieved the highest score—is not available via public API. It is limited to subscribers of Muse Code, Meta's integrated development environment. That is, the model ranking #2 in overall intelligence is precisely the one that the broader developer audience cannot access directly. The base Muse Wonder, which is available on the Meta Model API, scored 52 on the same index, which still places it among the world's best, but well below the top 2.
This reflects a deliberate strategy: Meta is prioritizing the model's integration as the engine of its own product ecosystem (Muse Code, the Muse app, future augmented reality experiences) before opening it as an API commodity. In contrast, Anthropic keeps Opus 5.5 widely available on the Anthropic API, while OpenAI keeps GPT-6 Astra within its own ecosystem. Each company is playing a slightly different version of the same game.
As for pricing, community sources suggest Muse Wonder would be priced at $4 per million input tokens and $20 per million output tokens, which would be exactly the same as Claude Opus 5.5. However, no official Meta pricing page confirms these figures, and the community reports themselves diverge on details—one listed cache at $0.02 per million, while another reported $0.20. The absence of an official pricing page raises questions about whether price parity is real or merely speculative.
The broader context of this race is that, as of October 2026, the frontier AI market has reached a point of unprecedented technical saturation. Google's Gemini 4 Argon leads DeepSWE 1.1 at 77.9%, Meta's Muse Spark 1.3 sits at 75.4%, OpenAI's GPT-6.1 Sol at 75.2%, and Claude Opus 5.5 at 74.2%. The margin between the top five models is only 3.7 percentage points—which means that, for many practical tasks, the choice among them depends more on pricing, API availability, and ecosystem than on genuine differences in capability.
What distinguishes the Muse Wonder Full launch is its timing: it arrived just as the dominant narrative was beginning to stabilize around a "Big Three" (OpenAI, Anthropic, Google) as the only real contenders for frontier AI. By presenting a model that competes directly on benchmarks, Meta is attempting to reinject its name into that conversation—and the bet is that its competitive edge lies not just in benchmark numbers, but in the ability to translate those models into useful agents for the user's everyday life.
The question that remains is whether a two-point gap on the Intelligence Index translates to real-world experience, or whether laboratory benchmarks have already reached the point of diminishing returns—where the next technical advance costs billions and produces differences imperceptible to the everyday user.
Sources: The Verge, Meta Research, Kie.ai
✓ Independent sources cross-checked and verified before publishing
