Muse Wonder Ties for Second in DualEntry's Accounting AI Benchmark
The accounting AI company says Meta's new model answered 83.2% of its benchmark tasks correctly, beating every OpenAI and Google model on the test but trailing
By Icarus2 min read
Meta’s Muse Wonder has tied for second place among 57 AI models on an accounting benchmark, the accounting AI company DualEntry said in a press release issued early Wednesday.
DualEntry says Muse Wonder answered 83.2% of tasks correctly on its Accounting AI Benchmark, up from the 61.4% scored by its predecessor, Muse Spark 1.1, on the same test. The result puts Wonder ahead of every OpenAI and Google model on the benchmark and level with Anthropic’s Claude Fable 5. Only xAI’s Grok 4.5 scored higher, at 84.2%.
The release, distributed over BusinessWire, adds detail: 97% accuracy on financial reporting and 92% on bank reconciliation, with 86%, 87%, and 77% across easy, medium, and hard tasks. The strongest OpenAI model on the test was GPT-5.5 Pro at 80.2%; the strongest Google model was Gemini 3.7 Flash at 82.2%.
These are DualEntry’s own figures, so some caution is in order. The company sells AI accounting tools, and its benchmark is synthetic rather than real corporate books: 101 constructed cases across eight areas (transaction classification, journal entries, financial reporting, bank reconciliation, accounts payable, accounts receivable, month-end close, and accounting knowledge), graded against fixed reference answers with an 80% pass mark, with each model running the full set multiple times to test consistency. DualEntry publishes the methodology and full results for all 57 models on its site.
Reporting: The Model’s Strong Suit
Where Wonder stood out, according to DualEntry, was reporting: the company said the model answered almost every reporting question correctly across repeated runs, including cash flow and financial statement calculations. Its misses clustered in the transaction work underneath — accounts receivable payment scenarios, month-end close journal entries, and edge cases in transaction classification.
“Muse Wonder reads financial statements as well as any model we’ve tested,” DualEntry co-founder Benedict Dohmen said in the release. “Where it still slips is the receivables and journal entries underneath them, which is exactly the work finance teams want an agent to take off their plate.”
Why It Matters
Meta is positioning Wonder for professional work, and the launch materials singled out finance alongside law and health. A strong showing in a finance benchmark two days after launch supports that pitch, even as Meta patches deployment bugs: Meta’s AI chief publicly apologized early Wednesday for the “ups and downs” of the model’s rollout while the team fixes inference bugs.
DualEntry’s caveat stands: finance is where agent mistakes cost real money, and transaction-level journal entries are exactly what a company would need an agent to get right before handing it the books. For now, Wonder reads the statements well and still needs supervision on the ledger lines underneath them.
