Which AI models should you test for document extraction?
Compare published reading and visual evidence, then test the candidates on the documents you actually handle.
Where to start today
GPT-6.1 Sol leads this published-evidence mix. Start a comparison here if you're reconsidering your setup. A higher index isn't proof that switching will improve your work.
Reading carries three quarters of this mix. Visual reasoning contributes the remaining quarter for scans and layouts. For text-only documents, remove visual reasoning in your saved comparison.
| Rank | Model | Index | Tests |
|---|---|---|---|
| 1 | GPT-6.1 Sol | 74.31 | 4 |
| 2 | GPT-6 Astra | 73.37 | 6 |
| 3 | Claude Fable 5 | 71.07 | 5 |
| 4 | Claude Fable 5.1 | 70.47 | 5 |
| 5 | Claude Opus 5.5 | 69.95 | 5 |
Leading open-weight option: Kimi K3. Downloadable weights don't imply an unrestricted licence or affordable hosting.
Adjust the mix and save it to watch your current model. No account needed. Get the weekly email for reports and confirmed releases; we skip quiet weeks.
Before you switch
Use a small set of your own documents with checked answers, including a difficult scan and a long file. Compare missing fields, wrong values and citations with your current model. Include retries and review time in the cost.
What this evidence misses
The published tests don't see your scans, layouts or edge cases. Reading scores don't establish citation accuracy or reliable output schemas. Some reading and visual tests overlap, so this mix can give those tasks extra weight.
Each capability needs at least two published tests to earn a rank. Coverage and test settings can differ between models; small score gaps may be noise. The comparison with your current model uses shared tests. Prices and production latency aren't included in this mix.
Published benchmark sources: Blueprint-Bench 2 · ExtractBench: document extraction · GDP.pdf (Surge AI) · LiveBench: Language + Data Analysis · Vision Evals (low effort) · Vision Evals: OCR & extraction (low effort)
What changed
Last report check: Oct 19, 2026. Only reports affecting this guide appear above.