DeepSeek’s V4.1 Flash 1017 got 82.2% of tasks right on the DualEntry Accounting AI Benchmark, down from 84.2% for the earlier DeepSeek V4.1 Flash. The update ties for tenth of 19 models, level with Zhipu AI’s GLM-5.3 and Anthropic’s Claude Opus 5.5.

Open models are close on accuracy but have not caught up. Zhipu AI’s GLM-5.3 Flash is the best open model at 88.1%, three points behind Google’s Gemini 4 Argon at 91.1%. Closed models also hold the top three scores on hard tasks and the top five on consistency.

Key results:

  • DeepSeek V4.1 Flash 1017: 82.2%, down from 84.2% for DeepSeek V4.1 Flash

  • Leader: Gemini 4 Argon (closed) at 91.1%. Best open model: GLM-5.3 Flash at 88.1%

  • Hard tasks: the top score, 82.1%, is shared by three closed models: Gemini 4 Argon, Grok 4.7 and Gemini 3.8 Flash

  • Consistency: Gemini 4 Argon passed 86.1% of tasks in all three runs. GLM-5.3 Flash passed 76.2%

  • Speed: GLM-5.3 Flash and GLM-5.3 averaged 57 and 62 seconds per task. Closed models in the top ten took 17 to 34 seconds

  • [DeepSeek V4.1 Flash 1017 consistency, cost per task and weakest category, from the benchmark team]

Consistency matters more than a single score for finance work. A model that books an entry correctly in one run and incorrectly in the next can’t be trusted to close the books without review. On that measure, the closed models from Google, Anthropic and Meta still lead.

“Open models are close enough on accuracy that the score gap is no longer the story,” said Benedict Dohmen, founder of DualEntry. “The gap is reliability. An agent working in the general ledger has to give the same answer every time it runs, and closed models are still ahead on that.”

The DualEntry Accounting AI Benchmark (DualBench100) tests models on 101 synthetic accounting cases across eight areas: transaction classification, journal entries, financial reporting, bank reconciliation, accounts payable, accounts receivable, month-end close and accounting knowledge. The cases split into 37 easy, 36 medium and 28 hard. Each case has a fixed reference answer and is scored by deterministic rules with an 80% pass mark. Every model runs the full set three times. A task counts as passed when the model passes it in at least two runs, and consistency is the share of tasks passed in all three. Results for all 19 models are at dualentry.com/benchmark.

About DualEntry

DualEntry is the agentic, AI-native ERP for mid-market finance teams. It brings the general ledger, accounts payable and receivable, revenue recognition, multi-entity consolidation and reporting into one platform, with AI agents that take on accounting work from data entry through close. Founded in 2024 and based in New York, DualEntry raised a $90 million Series A in 2025 from Lightspeed Venture Partners, Khosla Ventures and GV. Learn more at dualentry.com.

Media gallery

About The Author
Business Wire