GLM 5.2 Prepares a UK VAT Return Within 7 Pence of Human Bookkeepers for $2.73 of Tokens

Toot Blog (via Hacker News, 213 points)·medium signal

A benchmark on Vineyard Finance's real Q1 2026 books ran 59 transactions through GLM 5.2 in 68 minutes at $2.73 in raw token cost, landing the net VAT position (Box 5) just 7 pence — about 10 US cents — off the human-prepared ground truth. The author is explicit about the caveat that undercuts the headline: the humans also had to locate relevant invoices and reason about circumstances not inferable from the bank feed and receipts alone, so the model was handed a narrower task. Treat this as evidence about structured back-office extraction, not end-to-end bookkeeping replacement.