About GrindBench
GrindBench tests AI models on the ordinary paperwork people are paid to do, marks every answer against a checklist of what a careful manager would catch, and publishes what each model got right and what it cost.
Who runs it
GrindBench is independent of the labs whose models it tests. Every model is run the same way, through OpenRouter, with the same instructions and no special settings. No lab pays to be included, sees the scored tasks, or reviews results before they are published.
The marking is done by code wherever an answer can be checked exactly, and otherwise by a separate model from a lab none of the tested models come from, working down a fixed checklist.
What these numbers can't tell you
- They cover the work we tested, and nothing else. GrindBench-PubSec: atip redaction, procurement evaluation, performance analysis, bilingual equivalence. GrindBench-Banking: lending & mortgages, aml & kyc, deposits & payments, complaints & conduct, credit risk, capital & liquidity. A model's score here is a poor guide to work that isn't on this list.
- One attempt per task. Each model did each task once. Run it again and some answers would change.
- Every score has a range. We show a 95% range beside each score and put models whose ranges overlap in the same group. Inside a group, the order means little.
- GrindBench-Banking is a short run. 60 tasks per model, and only 10 for each kind of work. A score for a single kind of banking work can easily be 20 points off. The overall score is steadier.
- Canadian work, in English and French. The laws, forms and institutions are Canadian. Results may not carry to other countries.
- Cost is the model's bill only. Tokens at list price. It leaves out the person who checks every draft, which usually costs far more.
- The marker is automatic, and automatic markers make mistakes. We test ours before every round and publish how it did, and we audit our answer key when models agree against it. We have found errors that way.
- "Right" is a high bar. No deal-breaker mistakes and at least 80% of the checklist. A draft that fails may still save someone time; we don't measure that.
The data
Every number on this site comes from one file of published results: each model's share of work right, its range, and its cost, overall and for each kind of work.
| Benchmark | Models | Tasks each | Latest run |
|---|---|---|---|
| GrindBench-PubSec | 28 | 300 | 2026-10-29 |
| GrindBench-Banking | 18 | 60 | 2026-11-03 |
Visits are counted with Vercel Web Analytics: which pages are read and which site sent the reader. It sets no cookies and stores nothing that identifies you.
Use it freely with a link back. If you find an error, tell us: corrections are listed under What's new.