Muse Wonder tops TaxCalcBench but used web search to look up an answer

Muse Wonder ranks first on TaxCalcBench, a tax-filing benchmark, according to Alexandr Wang, who shares the result on X. Michael Bock reports that on one of the 50 tasks, the model used its web search tool to find the exact answer on GitHub, where the benchmark's code and data are open source.

Bock also says Muse Wonder costs almost $3 per return.