ModelVerdict

Adam Vítek · 2025/1045

CzechClean image13 fields

Document cs-p0025 · layout cs-L01 · 13 fields · public part of the dataset · Same invoice in English (US) → · Report an error in this example

Adam Vítek · 2025/1045

Verdict

Cheapest without an error: DeepSeek V4.1 Flash at $0.53 per 1,000 documents.

  • GPT-6 Sol is also error-free but 26.4× more expensive.
  • Gemma 4 31B is 3× cheaper, but: wrong value (Bank account).
  • Mistral Small 4 is 1.1× cheaper, but: invented value (Tax point).

Prompt given to every model

Z přiloženého dokladu vytěž údaje a vrať pouze JSON s těmito klíči:

supplier_name      název dodavatele (přesně jak je na dokladu, včetně diakritiky a právní formy)
supplier_id        IČO dodavatele
vat_id             DIČ dodavatele
invoice_number     číslo dokladu
payment_reference  variabilní symbol
issue_date         datum vystavení (YYYY-MM-DD)
tax_point_date     datum uskutečnění zdanitelného plnění, DUZP (YYYY-MM-DD)
due_date           datum splatnosti (YYYY-MM-DD)
net_by_rate        základ daně podle sazeb DPH, např. {"21": 18450.00, "12": 1200.00}
vat_by_rate        DPH podle sazeb, např. {"21": 3874.50}
total              celkem k úhradě (číslo, záporné u dobropisu)
currency           měna podle ISO 4217 (CZK, EUR, …)
bank_account       číslo účtu dodavatele (tvar předčíslí-číslo/kód banky) nebo IBAN

Pravidla:
- Údaje odběratele nevracej, jen dodavatele.
- Pokud údaj na dokladu není, vrať null. Nic si nedomýšlej.
- Pokud doklad neobsahuje rozpis DPH (neplátce, zálohová faktura), vrať net_by_rate a vat_by_rate jako null.
- Částky jako čísla s desetinnou tečkou, bez oddělovačů tisíců a bez měny.

How we score

Exact match per field after normalising dates, amounts and whitespace. Missing diacritics count as an error. Cost is the real token spend of this run, converted at the rate shown in the footer.

What each model extracted

FieldCorrect valueGemma 4 31B12 / 13 fieldsMistral Small 412 / 13 fieldsDeepSeek V4.1 Flash13 / 13 fieldsGPT-6 Luna13 / 13 fieldsGemini 3.5 Flash Lite13 / 13 fieldsQwen3.8 Flash12 / 13 fieldsGemini 3.8 Flash13 / 13 fieldsMistral Medium 3.512 / 13 fieldsClaude Sonnet 513 / 13 fieldsGPT-6 Sol13 / 13 fields
SupplierAdam VítekAdam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)Adam Vítek (correct)
Supplier ID (EIN, IČO…)6123147961231479 (correct)61231479 (correct)61231479 (correct)61231479 (correct)61231479 (correct)61231479 (correct)61231479 (correct)61231479 (correct)61231479 (correct)61231479 (correct)
VAT ID—— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)
Invoice number2025/10452025/1045 (correct)2025/1045 (correct)2025/1045 (correct)2025/1045 (correct)2025/1045 (correct)2025/1045 (correct)2025/1045 (correct)2025/1045 (correct)2025/1045 (correct)2025/1045 (correct)
Payment reference2025104520251045 (correct)20251045 (correct)20251045 (correct)20251045 (correct)20251045 (correct)20251045 (correct)20251045 (correct)20251045 (correct)20251045 (correct)20251045 (correct)
Invoice date2025-09-132025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)2025-09-13 (correct)
Tax point—— (correct) 2025-09-13invented value— (correct)— (correct)— (correct)— (correct)— (correct) 2025-10-13invented value— (correct)— (correct)
Due date2025-10-132025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)2025-10-13 (correct)
Net by tax rate—— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)
Tax by rate—— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)
Total253449.53253449.53 (correct)253449.53 (correct)253449.53 (correct)253449.53 (correct)253449.53 (correct)253449.53 (correct)253449.53 (correct)253449.53 (correct)253449.53 (correct)253449.53 (correct)
CurrencyCZKCZK (correct)CZK (correct)CZK (correct)CZK (correct)CZK (correct)CZK (correct)CZK (correct)CZK (correct)CZK (correct)CZK (correct)
Bank account392286-346812831/2010 39286-346812831/2010wrong value392286-346812831/2010 (correct)392286-346812831/2010 (correct)392286-346812831/2010 (correct)392286-346812831/2010 (correct) 392286–346812831/2010wrong value392286-346812831/2010 (correct)392286-346812831/2010 (correct)392286-346812831/2010 (correct)392286-346812831/2010 (correct)
Tokensinput / output704 / 2022,592 / 1731,746 / 4185,165 / 1821,516 / 2023,286 / 8721,516 / 5212,592 / 1735,317 / 1915,165 / 107
of which hidden reasoningbilled as output00267730694319000
of which from cacheinput billed at a discount4270614002,9440000
Cost of this documentthis run$0.0002$0.0005$0.0005$0.0007$0.0010$0.0012$0.0031$0.0052$0.0125$0.0140
Per 1,000 documentsfor the same kind of document$0.18$0.49$0.53$0.74$0.96$1.23$3.09$5.19$12.54$13.98
Latencythis run10.4 s1.9 s4.4 s2.9 s1.8 s18.6 s2.7 s3 s3.5 s3.3 s