RealRun / Guides
Short answer: In our test of 8 AI models on 79 receipts, bills and invoices, GPT-6.1 Sol scored highest (95.8 of 100) with no severe errors. GPT-6 Luna (91.4) also had no severe errors and was the cheapest. We don't recommend Grok 4.3 (69.2), which made up totals on unreadable receipts.
Headline score combines field, category and flag accuracy and how often a person was asked to check a bad document, minus 5 points per severe error. Average of 2-3 passes.
| Model | Score | Held-back 20 | Severe errors | Cost |
|---|---|---|---|---|
| GPT-6.1 Sol | 95.8 | 97.0 | none | medium |
| Gemini 4 Argon | 93.6 | 96.2 | 1 of 3 passes | medium |
| Claude Haiku | 92.8 | 95.5 | 1 of 3 passes | subscription |
| GPT-6 Luna | 91.4 | 93.2 | none | low |
| Claude Opus | 89.7 | 95.2 | both passes | subscription |
| Gemini 3.5 Flash-Lite | 85.4 | 87.6 | both passes | low |
| Grok 4.7 | 83.1 | 96.1 | both passes | high |
| Grok 4.3 | 69.2 | 89.1 | every pass | medium |
79 documents (59 main plus 20 held back), including 15 degraded photos, 5 receipts that don't add up and 6 fake invoices. Same frozen prompt for every model, scored automatically against an answer key with no AI judge. Full receipts results.
Want the exact prompt and setup? The Receipts AI Playbook has the word-for-word tested prompt, model picks by budget, and every miss explained.
Scores reflect our tests on the dates shown, not a guarantee. Claude models ran on a Claude subscription, so they have no metered API cost. cost_tier is relative within a task (low/medium/high) from our run spend; 'subscription' = not metered. Tested October 2026 by RealRun (Base Ledger IP LLC).