开源大模型 GLM 5.2 在英国中小企业季度增值税（VAT）申报测试中的表现评估显示，该模型能以低于人类会计师 1% 的成本处理复杂的会计合规工作 [1]。尽管存在少量非财务影响的分类错误，其生成的增值税申报表净额与真实值仅相差 7 便士（约 0.10 美元），近乎完美地完成了任务 [1]。
在具体的测试过程中，GLM 5.2 处理了 59 笔交易，耗时 68 分钟，原始 Token 成本仅为 2.73 美元 [1]。相比之下，人类会计师完成同类服务的典型费用约为每季度 750 至 2,100 英镑 [1]。
模型在评估中主要暴露出的错误是将价值约 13,300 美元的创始股份（10,000 英镑）误记为“资本账户”，而非正确的“未支付股份”类别 [1]。此外，该模型在总共 354 项评分标准中失败了 20 次，其中包含 14 次混淆了“零税率”与“免税”类别的情况 [1]。
Open-source large language model GLM 5.2 demonstrated performance nearly matching that of a human bookkeeper when evaluated on quarterly Value Added Tax (VAT) filing tasks for small and medium-sized enterprises in the UK [1]. The assessment revealed that the model could handle complex accounting compliance work with costs lower than those charged by humans, despite committing minor classification errors unrelated to financial impact [1]. In processing 59 transactions, GLM 5.2 required just 68 minutes of time and incurred an original token cost of only $2.73 USD [1]. The net amount generated in Box 5 of the VAT return produced by the model differed from the true value by merely 7 pence (approximately 10 US cents) [1].
While highly accurate, the model did encounter specific errors during testing; notably, it misclassified £10,000 GBP (roughly $13,300 USD) of founder shares as "Capital Account" instead of the correct category "Unpaid Shares" [1]. The system failed 20 out of 354 scoring criteria, with 14 of those failures involving a confusion between zero-rated and exempt categories [1]. In contrast, typical fees for human accountants performing similar services range from £750 to £2,100 GBP per quarter [1].