Summary
Theo argues that Meta has finally shipped a frontier-tier model with Muse Wonder, putting a lab other than OpenAI or Anthropic in the same class as Fable and Astra. After several days of testing, he says the model surprised him, and he points to benchmark results that place the top Wonder variant third overall on Artificial Analysis, ahead of every OpenAI model.
He starts with the criticisms. The release is split into two models, Muse Wonder, running at roughly 120 tokens per second, and the larger Muse Wonder Full, at about 50. Only the smaller one is available through the API with pricing, while the Full model is limited to a Muse Code subscription, which makes price-performance comparisons hard. Meta also left pricing out of its announcement. Memory is a central focus, with the model trained during post-training to understand how memory works.
The most notable findings concern coding and orchestration. In T3 Code, Muse and Claude Opus reviewed each other’s plans, and Opus often adopted Muse’s approach as the base. Theo found Muse a strong commander of sub-agents, writing clear delegation prompts and handling many parallel tasks, a skill he says labs outside OpenAI and Anthropic have typically struggled with.
📺 Source: Theo – t3․gg · Published October 13, 2026
🏷️ Format: Review
