Meta ships Muse Wonder, its first frontier AI model built to challenge OpenAI and Anthropic

Meta's Superintelligence Labs released two new AI models on October 12, 2026: Muse Wonder and Muse Wonder Full, the company’s first frontier-class models and its most serious bid yet to challenge OpenAI and Anthropic in the enterprise AI market.

The release follows Muse Spark, which debuted in April 2026 and landed with mixed reviews — it narrowed the gap with frontier models from OpenAI and Anthropic but failed to close it. Meta spent the intervening six months training the successor at its Prometheus data center in central Ohio, a facility purpose-built for large-scale AI workloads.

Two tiers, different access paths

Muse Wonder is available to developers immediately through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full — the more powerful of the two — is reserved for subscribers of Muse Code, Meta’s AI coding platform, which starts at $50 per month.

Both models will eventually power the consumer Muse personal assistant app, though Meta says the rollout begins with Muse subscribers first.

Benchmarks: close, but not ahead

On the Artificial Analysis Intelligence Index, Muse Wonder Full scores 56 and Muse Wonder scores 52. Claude Opus 5.5 leads at 58, and GPT-6 Astra sits around 53. By that measure, Wonder Full has nearly closed the gap with Anthropic’s top model and outpaces OpenAI’s — a significant improvement from Muse Spark’s position six months ago.

Coding is where Meta makes its strongest claim. On the SWE-Marathon benchmark, Muse Wonder Full scored 57.5, compared to Claude Opus 5.5’s 46.3 and GPT-6 Astra’s 42.5. Meta acknowledges these numbers were produced at maximum effort and cautions against direct comparisons without controlled conditions, but the gap is large enough to warrant attention from engineering teams evaluating AI coding assistants.

Active Memory and Active Focus

Meta highlights two capabilities that distinguish the Muse Wonder series from its predecessors. Active Memory enables the model to retain relevant context across conversations — a persistent working memory that carries forward what you’ve told it before. Active Focus allows the model to keep working toward a goal when tasks get difficult: trying different approaches, recovering from errors, and checking in before taking consequential actions rather than silently failing or giving up.

Both features are designed for agentic workflows — the kind of multi-step, autonomous task execution that enterprise buyers are increasingly demanding.

Meta’s commercial AI pivot

The release marks a genuine inflection point for Meta’s AI business. Muse Spark was essentially a research landmark — proof that Meta’s Superintelligence Labs under Alexandr Wang could build a frontier-class model. Muse Wonder is the commercial follow-through, with API pricing, enterprise coding tiers, and benchmark numbers positioned against specific competitors.

Meta’s 2026 capital expenditure plan of $115–135 billion is its bet that direct model revenue can become a meaningful second pillar alongside its advertising business. Muse Wonder is the product that has to start delivering on that promise.

As first reported by TechCrunch and confirmed by Meta’s official newsroom.