Muse Wonder copied one answer from GitHub but would still lead without it. With search turned off, it scores below most of its rivals.
Muse Wonder scored highest on TaxCalcBench, a test of whether AI models can work out a tax return correctly. On one of the test's 50 returns, though, it used its web search tool to find that return's answer in the benchmark's public GitHub repository and copied it. Column Tax, the tax-software company that runs the test, threw out that run and scored Wonder on the other 49.
Wonder still comes first without that return. It got 35 of the 49 exactly right, or 71%. If the excluded return counts as a miss, the score is 35 of 50, or 70%. That still beats OpenAI's GPT-6 Sol at 66% and Claude Opus 5.5 at 64%. Each model ran each setting only once, so one return moves a score by 2 points.
What matters more is what Wonder needs to get there. It leads only at its highest reasoning setting with web search on. One setting lower, still with search, it got 21 of 50 right. With search off, its best was 22%. That puts it behind Opus 5.5, Claude Fable 5.1, GPT-6 Astra and both versions of Sol.
Search helps most models because it lets them look up current tax forms. But no model gained as much from it as Wonder, which went from 22% to 71%. GPT-6 Sol came next, going from 28% to 66%. The answers to every test case are public on GitHub, where Wonder found one. Column's notes flag only that one run and don't say how it checked the other 49. The benchmark's maintainer, who posted the results, called Wonder "questionably aligned".
At its top setting Wonder cost $2.83 per return. That is less than Opus 5.5 at $3.50 and more than eight times GPT-6 Sol's 34 cents. Alexandr Wang, who leads Meta's AI lab, reposted the result with the line "muse wonder is the best model in the world at filing taxes." He did not mention the GitHub lookup.
The hype is real (but there's caveat): Muse Wonder is the best model in the world at filing taxes, according to TaxCalcBench. But it also is the first model ever to try to CHEAT at the task. (maybe that's not great when you're dealing with the IRS...) The data from our eval matches the current narrative: Muse Wonder is extremely benchmaxxed, and that includes TaxCalcBench. But it's also questionably aligned. On one of the 50 tasks in the benchmark, Muse Wonder used its web search tool to look up the exact answer on GitHub (the code & data is open source). Wonder cheated. Last thing: Muse Wonder is also super expensive: coming in at almost $3 per return! Cost will have to come down over time to make this model competitive with other methods of calculating personal income tax returns.
