GLM 5.2 is nearly as accurate as a human book keeper
toot-books.pages.dev
July 9, 2026
9 min read
🔥🔥🔥🔥🔥
59/100
Summary
GLM 5.2 achieves accuracy comparable to a human book-keeper in preparing quarterly VAT returns for small UK businesses. The cost of using GLM 5.2 is less than 1% of traditional accounting services.
Key Takeaways
- GLM 5.2 can prepare a quarterly VAT return for a UK SME with 99.99% accuracy, processing 59 transactions in 68 minutes at a cost of 2.73 USD.
- The model's output was only 7 pence off from the correct VAT return amount, demonstrating its high precision in financial tasks.
- GLM 5.2 utilized a command-line interface to input transaction data into accounting software and had access to the internet for specific operational queries.
- The benchmark was conducted with the model isolated from the ground truth to ensure the integrity of its performance evaluation.
Coverage & discussions
Hacker News · toot-books.pages.dev
Community discussions
Hacker News discussion ↗
Historical count: 5 · 5 saved in sample · Historical sample
Analysis ready
Community Sentiment
Positives
- This benchmark reveals that GLM 5.2 is impressively accurate, nearly matching human bookkeepers, which could redefine the accounting landscape.
- Automating menial tasks has been a game changer for many, showing the real-world potential of AI in streamlining accounting processes.
- The ability to hook up various data sources seamlessly with AI is a huge leap forward for efficiency in bookkeeping.
Concerns
- The liability concerns around AI making decisions in accounting are staggering — if an LLM commits fraud, who's accountable?
- Critics argue that benchmarks often fail to capture the nuanced realities of bookkeeping, raising doubts about the model's true capabilities.
- Humans are not exactly known for perfect recall, so comparing an LLM's accuracy to that of a human bookkeeper seems misleading.
