Meta's Muse Wonder Full tops OpenAI on legal and automation tests, but selected wins aren't overall leadership

Meta has a new model and a new set of claims. According to the reported results, Muse Wonder Full beat Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on automation and legal tasks. On 'Val', an index that measures professional performance, it reportedly scored 64.57% accuracy, ahead of every OpenAI model.

The tasks matter. Legal work and workflow automation are close to how companies actually use these models and pay for them. That makes these results more relevant than a chatbot popularity contest. The 64.57% figure also shows how much room is left: on this index, the best reported result still misses about a third of professional tasks.

Our short source leaves out a lot. It does not give the full benchmark table, results on harder coding or agentic tests, or who ran the evaluations. We have written before about leaderboards whose scores are 'optimised for voters', with Llama 4 as a clear case. That history means Meta's results need independent replication before anyone treats them as settled.

Our reading: beating the leaders on a few chosen tests is a claim about specific tasks, not about overall leadership. The test is whether Muse Wonder Full also does well on benchmarks that are hard to game, such as SWE-bench Pro and Terminal-Bench, which we track. If it does, Meta is back at the frontier and the market gets one more serious competitor. More competition at the top lowers prices and spreads access, which supports our long-term view that AI drives abundance. If it doesn't, this is marketing, and we will grade it on the evidence.