Meta’s Muse Wonder ties for second on accounting benchmark

Meta describes Muse Wonder and Muse Wonder Full as its most capable models to date, with state-of-the-art performance on personal and professional agentic tasks.[1] The company says Active Memory carries important personal context across conversations and retrieves relevant details for a task.[1] Meta also says Muse Wonder can continue toward challenging goals, try alternatives, work through blockers and adapt to new information.[1] It says the model asks for clarification when needed and checks in before important actions.[1]

Meta says Muse Wonder is available through its Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers.[1] The company says Muse Wonder will soon power experiences in Muse, with rollout beginning for subscribers.[1] Meta says it trained both models at its Prometheus data center.[1] It also says the models include safeguards against chemical, biological, radiological and nuclear risks, as well as cyber misuse.[1]

DualEntry’s benchmark covered 101 synthetic accounting cases across eight areas.[2] Muse Wonder scored 83.2% overall and tied for second among 57 models; Grok 4.5 scored 84.2%.[2] Muse Wonder’s result was 21.8 percentage points higher than Muse Spark 1.1’s.[3] The benchmark uses fixed reference answers, deterministic scoring and an 80% pass mark, and runs models several times to measure consistency.[2] Those design choices provide a repeatable comparison, but the result is specific to the benchmark’s cases and scoring method.[2]

Muse Wonder scored 97% on financial reporting and 92% on bank reconciliation; its scores were 86% on easy tasks, 87% on medium tasks and 77% on hard tasks.[2] DualEntry said its reporting answers were nearly all correct across repeated runs, including cash-flow and financial-statement calculations.[2] The benchmark also found errors in some accounts-receivable payment scenarios and inconsistency on month-end-close journal entries and transaction-classification edge cases.[2] DualEntry co-founder Benedict Dohmen said the remaining weaknesses involve work finance teams want agents to handle.[2]

Artificial Analysis reported Intelligence Index scores of 52 for Muse Wonder and 56 for Muse Wonder Full at maximum reasoning effort; Muse Spark 1.3 scored 48.[4] Muse Wonder Full placed in the top three on six of the ten Intelligence Index evaluations and scored 68 on the Coding Agents Index at maximum effort in Muse Code.[4] Artificial Analysis listed Muse Wonder at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, while Muse Wonder Full is not priced per token.[4] Both models have one-million-token context windows.[4]