Meta launches Muse Wonder models claiming frontier parity, but the strongest version stays locked behind subscriptions

Led by Alexandr Wang, MSL built the models on Meta's $135 billion 2026 infrastructure expansion. They target coding, legal analysis and financial reasoning. Vals AI reported that Muse Wonder Full (Max) debuted at #5 on its index, surpassing every OpenAI model. Bloomberg reported that Zuckerberg admitted Meta lacked AI certainty on its $200 billion bet.

The benchmark picture is mixed. Wonder posts DeepSWE 78.8 versus Opus's 74.2 and SWE-Marathon 53.1 versus 46.3, but ProgramBench shows Opus leading 70.5 to 50.0, and some developers call the results 'benchmaxxing'. The API Wonder scores 52 on the Artificial Analysis index, behind Opus 5.5's 58, at identical $4/$20 pricing. The stronger Wonder Full (56) is subscription-only.

The strategic shift is that Meta, once the champion of open-weight Llama, now keeps its best model closed. Developers are debating the U-turn, and some report the Muse Code sandbox 'never worked across Mac or Linux'.

Watch for API access to Wonder Full, pricing changes, and Meta's response to Gemini 4 Argon overtaking it on Vals a day later.