Muse Wonder

Meta Platforms released Muse Wonder and Muse Wonder Full on October 12, 2026 from its Superintelligence Labs (MSL), targeting personal and professional agentic workloads. The launch follows a 14-month AI division overhaul led by CEO Mark Zuckerberg and Chief AI Officer Alexandr Wang, who joined Meta in June 2025. Muse Wonder, internally codenamed "Watermelon," matches top rivals on coding, financial analysis, and legal reasoning tasks. Wang emphasized the team's differentiated focus on making the personal-agent use case work robustly across both consumer and developer surfaces. Meta committed billions of dollars toward AI infrastructure to support its roadmap. Wonder is available via Meta Model API and Muse Code, while Full is reserved for Muse Code subscribers. Microsoft President Brad Smith separately urged strict global AI regulation in the same news cycle.

On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full scored 56 at maximum reasoning effort, ranking second behind only Anthropic Claude Opus 5.5. Standard Muse Wonder scored 52, an 8-point jump over predecessor Muse Spark 1.3 which scored 48. Wonder Full also placed in the top three on six of the ten underlying evaluations, with strongest showings in agentic and coding benchmarks. Muse Wonder and Muse Wonder Full were released on October 12, 2026 via the Meta Model API and Muse Code, with Full limited to subscribers. Both models feature Active Memory and Active Focus for agent reliability, and Wonder is priced at $4 per million input tokens plus $20 per million output tokens. Meta claims the series beats Opus 5.5, GPT-6 Astra, and Claude Fable 5.1 on coding, legal, and automation tests.

Meta's research blog announced Muse Wonder and Muse Wonder Full on October 12, 2026 as the company's most capable frontier models for personal and professional agentic tasks. Both were trained at Meta's gigawatt-scale Prometheus data center, building on the Muse Spark scaling approach across pretraining and reinforcement learning. The series combines Active Memory and Active Focus capabilities for long-horizon agent reliability. Muse Wonder is available today through the Meta Model API and Muse Code, while Muse Wonder Full is reserved for Muse Code subscribers and not offered via the API. Meta states Wonder achieves state-of-the-art results on personal and professional agent benchmarks, plus memory and coding evaluations. Future plans include rolling Wonder into the Muse app for subscribers and artifact generation, with specialized derivative models planned for everyone using Muse.

Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026 as its most advanced models for personal and professional tasks such as navigating insurance claims, interpreting leases, and building applications. Wonder became available the same day through the Meta Model API and Muse Code, while Full is restricted to Muse Code subscribers. Active Memory retains context across conversations and Active Focus continues work on difficult tasks. Meta trained the Wonder models to avoid sharing private information without permission and to decline harmful requests, following its Superintelligence Scaling Framework. CEO Mark Zuckerberg told the Financial Times that Meta has reached the AI frontier, rejected calls to slow AI development, and noted Meta is already working on the next step up. Shares traded 3.7% higher on Monday following the announcement.

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, 2026 as its first frontier-class models, with Wonder available through the Meta Model API and Muse Code, and Full restricted to Muse Code subscribers. Wonder is priced at $4 per million input tokens and $20 per million output tokens, putting it in the same range as GPT-6 Sol and Opus 5.5. Both were trained at the gigawatt-scale Prometheus data center. On SWE-Marathon, which requires building whole projects like a C compiler over 2 to 10 hours, Muse Wonder Full scores 57.5 versus 46.3 for Opus 5.5, the widest published gap. On DeepSWE 1.1, Full leads 79.0 to 74.2 in third-party runs, and Full beats Opus 5.5 on 8 of 14 benchmarks per Meta's claims. Active Memory carries context across conversations while Active Focus keeps agents working through long tasks.

Meta released Muse Wonder on October 12, 2026 as its largest model, priced at $4 per million input tokens with $0.20 cached input and $20 per million output on the Meta Model API, matching Anthropic Claude Opus 5.5 pricing. The model offers a 1M-token context window. Wonder leads on long-running workloads, topping SWE-Marathon, DeepSWE, MemoryBench and AutomationBench in Meta's published benchmark table. On Meta's own 14-row benchmark comparison, the Wonder family leads 7 rows versus 6 for Claude Opus 5.5 and 1 for GPT-6 Astra. Plain Muse Wonder beats Opus 5.5 on 5 of 14 benchmarks, with Muse Wonder Full available exclusively to Muse Code subscribers rather than the API. A fuller safety report is expected later under Meta's Superintelligence Scaling Framework.