Meta Muse Wonder Review: Meta Reclaims the Frontier
Meta Superintelligence Labs releases its first frontier-class models, with Wonder Full outperforming GPT-6 Astra and Claude Opus 5.5 on key benchmarks.
Meta's Long-Awaited Comeback
On October 12, 2026, Mark Zuckerberg announced Muse Wonder and Muse Wonder Full, the first frontier-class models from Meta Superintelligence Labs (MSL). The release marks Meta's return to the top tier of AI model development after more than a year of setbacks following the underwhelming Llama 4 launch in April 2025.
Both models support a 1 million token context window and target coding, long-horizon agent tasks, and professional knowledge work. Wonder Full is the higher-capability variant, while the standard Wonder is positioned for broader API access. The release follows a massive organizational rebuild that included a significant investment in Scale AI, the recruitment of founder Alexandr Wang, and aggressive hiring from OpenAI and Google DeepMind with compensation packages reportedly reaching hundreds of millions of dollars.
Benchmark Results
Meta published extensive internal benchmarks alongside independent third-party evaluations. On the Personal Superintelligence Index, a composite metric developed by Meta covering professional work, automation workflows, and web operations, Wonder Full scored 59.1%, ahead of Claude Opus 5.5 at 55.2% and GPT-6 Astra at 54.5%. Even the standard Wonder achieved 54.3%, surpassing Claude Fable 5.1.
In AutomationBench-AA, which measures end-to-end business process automation, Wonder Full reached 45.4% compared to Astra's 41.6% and Opus 5.5's 41.2%. On MemoryBench, testing long-conversation recall with distractions and information updates, standard Wonder scored 85.6% and Wonder Full 85.5%, both ahead of Opus 5.5 at 81.5% and Astra at 81.4%.
Coding benchmarks showed particularly strong results. On DeepSWE 1.1, measuring long-horizon software engineering, Wonder Full scored 79.0% versus Astra's 74.1% and Opus 5.5's 74.2%. On SWE-Marathon, which requires sustained code maintenance across project lifecycles, Wonder Full achieved 57.5% success while Astra managed 42.5% and Opus 5.5 reached 46.3%.
Independent Verification
The third-party Artificial Analysis Intelligence Index provides an external reference point. Wonder Full scored 56 points, ahead of GPT-6 Astra at 53 but trailing Claude Opus 5.5 at 58. This suggests Meta's internal benchmarks, while directionally accurate, may paint a slightly more favorable picture than independent evaluations.
Meta also demonstrated that Wonder Full performs better on unfamiliar tasks than predicted by scaling trends from its predecessor Spark 1.3, suggesting stronger generalization from the same post-training investment. The models were trained at Meta's gigawatt-scale Prometheus data center with expanded pretraining and reinforcement learning phases.
Pricing and Access
Standard Muse Wonder is available through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, with cached input at $0.2. This matches Claude Opus 5.5's standard pricing and undercuts GPT-6 Astra's $10/$50 rate. Meta emphasizes that Wonder does not charge extra for long context usage.
Wonder Full is not available as a standalone API purchase. It is restricted to Muse Code's Power Usage ($50/month) and Maximum Usage ($200/month) subscription tiers, with the latter offering 80 times the usage quota of the base Everyday Usage plan.
Product Challenges Remain
Model capability is only one component of market success. According to a Wall Street Journal report on October 12, 2026, the Muse application has accumulated nearly 9 million downloads but growth is slowing and user retention trails ChatGPT significantly. Zuckerberg has spent 14 months rebuilding Meta's AI organization to produce competitive models. The next challenge is proving that Wonder can reinvigorate the Muse product and capture share in the consumer agent market. Meta faces stiff competition not only from OpenAI and Anthropic but also from a growing field of open-source alternatives and specialized vertical agents that threaten to fragment the market. Success will depend on whether Wonder can deliver consistently superior reasoning and coding performance in real-world tasks rather than just benchmark leaderboards.
Sources: 36氪, Apidog, ROO3, Meta Newsroom, Meta Newsroom