Hacker News front pageJul 10, 2026, 02:29 UTC+8

GLM 5.2 prepares a UK SME's quarterly VAT return at near-human accuracy for under 1% of the cost

GLM 5.2 is nearly as accurate as a human book keeper

Vineyard Finance tested GLM 5.2 on a real quarter of UK VAT bookkeeping: 59 transactions processed in 68 minutes at a raw token cost of $2.73. The final VAT return's Box 5 was off by only 7 pence. The model operated accounting software via CLI and was scored on 6 criteria per transaction; it failed 20 of 354 checks. The one serious error was misclassifying founder shares as 'Capital Account' instead of 'Unpaid Shares', which carries legal risk. The other 19 mistakes had no financial impact but are errors a skilled accountant wouldn't make. The model had internet access and two user notes; no cheating was detected. The post does not disclose the exact quantization, but believes it is FP16 or FP8.

Why it matters: A real business-scenario agent test, not marketing fluff. GLM 5.2 operated accounting software via CLI to complete a full quarter's VAT filing: 59 transactions, 68 minutes, $2.73 cost, Box 5 net off by 7 pence. Has scoring criteria and an error list — information density is so...

Read the original ↗Export Markdown

Take

I clicked because this is a real bookkeeping run, not a benchmark score. Vineyard Finance fed GLM 5.2 their actual Q1 2026 books: 59 transactions, processed via CLI into accounting software in 68 minutes, raw token cost $2.73. The final VAT return's Box 5 was off by 7 pence. That's the headline.

The catch: 20 of 354 checks failed. The one serious error was misclassifying founder shares as 'Capital Account' instead of 'Unpaid Shares' — that carries legal risk under UK company law. The other 19 mistakes had no financial impact but are errors a skilled accountant wouldn't make.

I'd discount this a bit. The model's job was narrower than a human's: no hunting through mailboxes for invoices, no reasoning about real-world context beyond two user notes. The model also knew it was being tested — its reasoning trace literally says 'the task is testing whether I get VAT right.' Quantization isn't confirmed; the post guesses FP16 or FP8.

The useful bit isn't accuracy, it's cost structure. Outsourcing a VAT return runs £750–2,100 per quarter. This run cost $2.73. But that one legally risky error means you wouldn't let it file unsupervised yet.

Source: Hacker News front page · toot-books.pages.dev
Published: Jul 10, 2026, 02:29 UTC+8 · Jul 10, 2026