Open-source AI models are narrowing the performance gap with proprietary systems in accounting tasks, according to the latest DualEntry Accounting AI Benchmark. The fourth-best score was recorded by Moonshot AI’s Kimi K3.1 Pro, which achieved an accuracy rate of 86.1%. In addition, Kimi beat all of Anthropic, OpenAI, and Meta’s models.
At the same time, Zhipu AI’s GLM-5.3 Flash also secured 88.1%, coming in second after Grok 4.7. It follows that the current top four is occupied by two open-source AI models.
Open-Source AI Models Deliver Competitive Results
Gemini 4 Argon by Google is the leading benchmark with an accuracy of 91.1%. Nevertheless, the top-performing open model lags behind its leader by merely three percentage points. The average score for seven open models was 83.9%, while 13 closed models got an average score of 81.3%.
In addition to this, Kimi K3.1 Pro outdid its predecessor, Kimi K3 from Moonshot, with 84.2%. Open models occupied four of the top seven spots in the leaderboard. It is noteworthy to mention the high performance of the models on particular accounting tasks. Thus, Kimi K3.1 Pro succeeded with 100% in the bank reconciliation test, while Gemini 4 Argon showed 75%.
Cost efficiency also distinguished open-source AI models. DeepSeek V4.1 Flash achieved 84.2% accuracy at just $0.003 per task. In comparison, Claude Fable 5.1 scored 85.1% at $1.18 per task.
Reliability Remains a Challenge
Even with these improvements, closed models still prove more consistent. The three closed models have the best scores in challenging tasks, attaining an 82.1% score. Furthermore, out of the five models that successfully completed individual tasks in all three tests, all were closed models, with Gemini 4 Argon scoring an 86.1%.
Benedict Dohmen of DualEntry explained how these findings could influence financial teams’ AI decisions.
“Finance teams have assumed that choosing an open model meant trading accuracy for cost and control,” said Benedict Dohmen of DualEntry. “On accounting work, that trade is mostly gone. What’s left is consistency, and that is the gap to watch.”
Inside the DualEntry Accounting AI Benchmark
The DualEntry Accounting AI Benchmark, or DualBench100, is an assessment of AI models using 101 artificial accounting case studies. The eight categories include transaction classification, journal entries, accounting, and bank reconciliation.
The standard will examine accounts payable, accounts receivable, month-end close, and accounting skills as well. This is done using a dataset that consists of 37 simple problems, 36 moderate problems, and 28 hard problems. There is a consistent reference solution and deterministic scoring scheme used for each problem. The minimum score required to pass is 80%. In addition, all models go through the full test set three times.
Explore Finance Tech News for the latest innovations in financial technologies and expert insights shaping the future of digital finance!
News Source: Businesswire.com
