Meta's Muse Wonder beats Claude Fable 5.1

Meta's return to the frontier follows a costly restructuring. Zuckerberg overhauled the company's AI research last year, sinking billions into the recently formed Meta Superintelligence Labs and hiring former Scale AI CEO Alexandr Wang to lead it. That effort was sometimes rocky, with some researchers not sticking around for long.

The company's earlier Llama models were the first frontier open-weight offerings from a big lab, but Meta fell behind after several disappointing versions, as Anthropic, OpenAI, and Google rapidly introduced more capable models. Meta announced the first model from its new lab, Muse Spark, in April of this year, and has since released a small open-weight model called Muse Glimmer. It did not disclose how the new Muse Wonder models were built or trained.

Alongside the benchmark claims, Meta says it trained Muse Wonder to refuse dangerous outputs and scores well on benchmarks measuring resilience to such prompts. The company did not say whether agents powered by Muse Wonder have misbehaved in testing or what it has done to mitigate that specifically. Zuckerberg said last month that the release had been delayed for safety work, and Meta revealed a framework earlier this month for scaling capabilities safely.