RealRun / Results
Tested on 112 bank exports, bill histories and statements (main + fresh, held-back sets) across 8 AI models, 2+ passes each: 1,992 graded answers.
Short answer: four models finished in a near tie: GPT-6 Luna, Gemini 4 Argon, GPT-6.1 Sol and Grok 4.7. Treat the gaps between them as noise. GPT-6 Luna was by far the cheapest ($0.66 per 1,000 documents); Gemini 4 Argon and GPT-6.1 Sol had no severe errors. We don't recommend Grok 4.3 or Gemini 3.5 Flash-Lite.
How the overall score (0-100) works: every run (one model, one pass, one part, main or held-back set) gets a weighted accuracy score, minus 5 points for each severe error in that run, floored at 0. A part's score averages the main and held-back runs within each pass, then averages the passes. Overall is the plain average of the three parts. The severe-error column totals every pass and set, so a counted error lowers the overall by much less than 5 points.
| Model | Overall | Categorize | Bill problems | Statements | Held-back | Severe errors | Cost / 1,000 docs |
|---|---|---|---|---|---|---|---|
| GPT-6 Luna | 96.3 | 96.3 | 96.7 | 95.9 | 95.6 | None | $0.66 |
| Gemini 4 Argon | 96.1 | 95.5 | 99.0 | 93.9 | 96.3 | None | $16.32 |
| GPT-6.1 Sol | 96.0 | 95.8 | 98.5 | 93.7 | 96.8 | None | $11.81 |
| Grok 4.7 | 95.6 | 94.0 | 98.5 | 94.4 | 95.4 | 1 | $34.91 |
| Claude Opus | 93.2 | 89.5 | 99.0 | 91.0 | 93.0 | None | Claude plan |
| Claude Haiku | 91.7 | 91.9 | 96.0 | 87.0 | 91.8 | None | Claude plan |
| Grok 4.3 ⚠ | 89.7 | 90.5 | 99.5 | 78.9 | 79.8 | 13 | $6.43 |
| Gemini 3.5 Flash-Lite ⚠ | 84.5 | 85.1 | 93.2 | 75.2 | 84.7 | 23 | $3.04 |
Severe errors are counted over every pass. Claude models ran on a Claude plan, so we have no per-document cost for them.
⚠ Not recommended: Grok 4.3. 13 severe errors; fell to 52-72 on fresh statements and got the export total right 17 of 80 times.
⚠ Not recommended: Gemini 3.5 Flash-Lite. 23 severe errors; followed a hidden 'say this reconciles' instruction in all 3 passes.
What you get in the $29 playbook (instant PDF download, 79 pages): all three word-for-word tested prompts, model picks with cost per 1,000 documents, every miss explained item by item, safety rules and setup steps (untested parts marked UNTESTED).
Know which AI keeps your books straight before you trust it with real money.
For you: catch the duplicate charge, the price hike and the scam bill some models wave right through.
For your clients: show them which models catch scam bills and which miss them, and ship the checks that stop bad numbers from reaching their books.
Tested on 112 bank exports, bills and statements with 1,992 graded answers. GPT-6 Luna, Gemini 4 Argon, GPT-6.1 Sol and Grok 4.7 finished in a near tie at the top.
Buy the finance playbook — $29 (founding price)
Or get all four playbooks for $79 (save $37 vs $116).
Data credits: transaction data adapted from DoDataThings (MIT) and AgamiAI (Apache-2.0).
Also see: receipts results, inbox sorting and meeting scheduling.