Gemma 4 31B vs GPT-6 Sol

Same documents, same prompt: where each model is right, what it really costs and how fast it answers.

At a glance

Task · documents	English (US)	Czech	German	Spanish	French
Invoice extraction	97.6% / 99.9%	96.1% / 99.9%	94.9% / 100.0%	95.4% / 99.2%	93.6% / 99.9%
Email triage	96.5% / 99.0%	97.1% / 99.0%	95.1% / 98.2%	—	—
Personal data detection	99.9% / 99.7%	99.8% / 99.7%	99.9% / 99.7%	—	—
Contract clauses	99.6% / 100.0%	100.0% / 100.0%	94.4% / 100.0%	—	—

Correct fields: Gemma 4 31B / GPT-6 Sol; the better one in bold. Click a cell to see the details.

Invoice extraction · English (US) (150 documents)

GPT-6 Sol is more accurate by 2.3 pp. Gemma 4 31B is 57.6× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	97.6%	99.9%
Error-free documents	81%	99%
Per 1,000 documents	$0.17	$9.65
To fix / 1,000	187	7
Speed (median)	8.9 sec	4.1 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Supplier	99.3%	100.0%
Supplier ID (EIN, IČO…)	98.0%	100.0%
Invoice number	100.0%	100.0%
Payment reference	99.3%	100.0%
Invoice date	98.7%	100.0%
Due date	98.7%	100.0%
Net by tax rate	95.3%	100.0%
Tax by rate	98.7%	100.0%
Total	100.0%	100.0%
Currency	100.0%	100.0%
Bank account	86.0%	99.3%

Documents where they differ

Harmon Logistics LLC · INV256017 · Scan · Gemma 4 31B 8/11, GPT-6 Sol 11/11
Austin Landscaping Inc. · INV-17836 · Scan · Gemma 4 31B 9/11, GPT-6 Sol 11/11
Gomez & Sons, LLC · 7855 · Scan · Gemma 4 31B 9/11, GPT-6 Sol 11/11
Summit Services Inc. · INV-15230 · Scan · Gemma 4 31B 10/11, GPT-6 Sol 11/11
Jennifer Moore · 5113 · Scan · Gemma 4 31B 10/11, GPT-6 Sol 11/11
Mason-Nelson Group LLC · 2025-3952 · Photo · Gemma 4 31B 10/11, GPT-6 Sol 11/11
Sarah Villanueva · 2025-7876 · Clean image · Gemma 4 31B 10/11, GPT-6 Sol 11/11
Johnson Electric LLC · INV251373 · Photo · Gemma 4 31B 10/11, GPT-6 Sol 11/11

Invoice extraction · Czech (150 documents)

GPT-6 Sol is more accurate by 3.8 pp. Gemma 4 31B is 47.8× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	96.1%	99.9%
Error-free documents	67%	99%
Per 1,000 documents	$0.21	$10.12
To fix / 1,000	333	13
Speed (median)	10.4 sec	3.9 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Supplier	96.7%	99.3%
Supplier ID (EIN, IČO…)	93.3%	100.0%
VAT ID	96.7%	100.0%
Invoice number	100.0%	100.0%
Payment reference	96.0%	99.3%
Invoice date	96.0%	100.0%
Tax point	95.3%	100.0%
Due date	94.0%	100.0%
Net by tax rate	98.0%	100.0%
Tax by rate	97.3%	100.0%
Total	99.3%	100.0%
Currency	100.0%	100.0%
Bank account	86.7%	100.0%

Documents where they differ

Jiřina Švecová · 25401227 · Clean image · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Bednář spol. s r.o. · 20250610 · Clean image · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Stejskal, v.o.s. · FV2600175 · Photo · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Ekolog s.r.o. · FV2600132 · Scan · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Hruška a Šťastný s.r.o. · 2025/0324 · Clean image · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Moravská energetická spol. s r.o. · 25101373 · Scan · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Michaela Strnadová · 25600363 · Scan · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Sedlák s.r.o. · 20250158 · Photo · Gemma 4 31B 12/13, GPT-6 Sol 13/13

Invoice extraction · German (150 documents)

GPT-6 Sol is more accurate by 5.1 pp. Gemma 4 31B is 53.5× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	94.9%	100.0%
Error-free documents	64%	99%
Per 1,000 documents	$0.20	$10.65
To fix / 1,000	360	7
Speed (median)	7.7 sec	4.9 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Supplier	98.7%	100.0%
Supplier ID (EIN, IČO…)	99.3%	100.0%
VAT ID	94.0%	100.0%
Invoice number	99.3%	100.0%
Payment reference	98.7%	100.0%
Invoice date	87.3%	100.0%
Tax point	92.0%	100.0%
Due date	90.7%	100.0%
Net by tax rate	94.7%	100.0%
Tax by rate	97.3%	100.0%
Total	99.3%	100.0%
Currency	99.3%	100.0%
Bank account	82.7%	99.3%

Documents where they differ

Spedition Müller GmbH · 2026-1446 · Scan · Gemma 4 31B 8/13, GPT-6 Sol 13/13
Wesermetall GmbH · RE2501375 · Scan · Gemma 4 31B 9/13, GPT-6 Sol 13/13
Zimmermann & Schulz GmbH · RE11274 · Scan · Gemma 4 31B 10/13, GPT-6 Sol 13/13
Dachdeckerei Fischer GmbH · 2025/0014 · Scan · Gemma 4 31B 10/13, GPT-6 Sol 13/13
Schmidt & Bauer GmbH & Co. KG · RE-2025-01397 · Scan · Gemma 4 31B 10/13, GPT-6 Sol 13/13
Richter GmbH & Co. KG · 2025-0509 · Clean image · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Nordtec GmbH & Co. KG · RE-2025-00909 · Scan · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Berger UG (haftungsbeschränkt) · GS10609 · Clean image · Gemma 4 31B 11/13, GPT-6 Sol 13/13

Invoice extraction · Spanish (87 documents)

GPT-6 Sol is more accurate by 3.8 pp. Gemma 4 31B is 52.8× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	95.4%	99.2%
Error-free documents	54%	89%
Per 1,000 documents	$0.20	$10.76
To fix / 1,000	460	110
Speed (median)	8.7 sec	5.7 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Supplier	95.4%	98.0%
Supplier ID (EIN, IČO…)	93.1%	100.0%
VAT ID	98.9%	100.0%
Invoice number	100.0%	100.0%
Payment reference	95.4%	100.0%
Invoice date	98.9%	100.0%
Tax point	100.0%	100.0%
Due date	97.7%	100.0%
Net by tax rate	81.6%	92.0%
Tax by rate	100.0%	100.0%
Total	100.0%	100.0%
Currency	100.0%	100.0%
Bank account	79.3%	99.0%

Documents where they differ

Coloma Servicios Informáticos S.L.U. · PRO-2025-0066 · Scan · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Talavera y Aller Asociados, S.L. · 2025/0610 · Scan · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Distribuciones Mayo Santana, S.L.U. · A-2025-0563 · Clean image · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Talleres Rozas Zapata S.L. · FV25-01486 · Clean image · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Construcciones del Norte S.L.U. · PF25/0020 · Clean image · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Centro de Formación del Sur S.L.U. · F2025-1424 · Scan · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Arranz e Hijos S.L. · F2026-0940 · Clean image · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Goñi Logística S.L. · R-2025/1460 · Scan · Gemma 4 31B 12/13, GPT-6 Sol 13/13

Invoice extraction · French (87 documents)

GPT-6 Sol is more accurate by 6.4 pp. Gemma 4 31B is 49.8× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	93.6%	99.9%
Error-free documents	51%	99%
Per 1,000 documents	$0.22	$11.05
To fix / 1,000	494	10
Speed (median)	7.2 sec	5.9 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Supplier	96.6%	100.0%
Supplier ID (EIN, IČO…)	77.0%	100.0%
VAT ID	95.4%	100.0%
Invoice number	97.7%	100.0%
Payment reference	96.6%	99.0%
Invoice date	96.6%	100.0%
Tax point	94.3%	100.0%
Due date	96.6%	100.0%
Net by tax rate	93.1%	100.0%
Tax by rate	100.0%	100.0%
Total	100.0%	100.0%
Currency	98.9%	100.0%
Bank account	73.6%	100.0%

Documents where they differ

Lefèvre & Fils SARL · FAC-2026-00198 · Photo · Gemma 4 31B 8/13, GPT-6 Sol 13/13
Guérin & Fils EURL · 2026-05-688 · Clean image · Gemma 4 31B 10/13, GPT-6 Sol 13/13
SA Établissements Gauthier · F2025-1451 · Photo · Gemma 4 31B 10/13, GPT-6 Sol 13/13
Nettoyage Christine Rivière EI · 2025-12-1174 · Clean image · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Établissements Bonnet EURL · AV-10599 · Scan · Gemma 4 31B 11/13, GPT-6 Sol 13/13
SARL Garcia Informatique · F2025-0111 · Clean image · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Bernard et Associés SASU · FAC-2025-01115 · Scan · Gemma 4 31B 11/13, GPT-6 Sol 13/13
Carrosserie Gauthier SASU · F-11079 · Scan · Gemma 4 31B 12/13, GPT-6 Sol 13/13

Email triage · English (US) (150 documents)

GPT-6 Sol is more accurate by 2.5 pp. Gemma 4 31B is 20.2× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	96.5%	99.0%
Error-free documents	72%	92%
Per 1,000 documents	$0.12	$2.44
To fix / 1,000	280	80
Speed (median)	2.2 sec	2.8 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Category	100.0%	100.0%
Priority	85.3%	98.0%
Sentiment	83.3%	93.3%
Customer name	100.0%	100.0%
Order number	100.0%	100.0%
Invoice number	100.0%	100.0%
Amount	100.0%	100.0%
Currency	100.0%	100.0%
Deadline	100.0%	100.0%

Documents where they differ

Daniel Stewart · Email · Gemma 4 31B 7/9, GPT-6 Sol 9/9
Allison Floyd · Email · Gemma 4 31B 7/9, GPT-6 Sol 9/9
Sheila Evans · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Gene Everett · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Laurie Knight · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Jason Davis · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Christine Copeland · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Angela Harmon · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9

Email triage · Czech (150 documents)

GPT-6 Sol is more accurate by 1.9 pp. Gemma 4 31B is 20.1× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	97.1%	99.0%
Error-free documents	75%	91%
Per 1,000 documents	$0.16	$3.19
To fix / 1,000	253	93
Speed (median)	2.5 sec	2.8 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Category	100.0%	100.0%
Priority	85.3%	98.0%
Sentiment	88.7%	92.7%
Customer name	100.0%	100.0%
Order number	100.0%	100.0%
Invoice number	100.0%	100.0%
Amount	100.0%	100.0%
Currency	100.0%	100.0%
Deadline	100.0%	100.0%

Documents where they differ

Alois Sedlák · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Lucie Kašparová · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Natálie Kratochvílová · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Šimon Beran · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Magdalena Sedláčková · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Ctibor Dušek · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Vanda Vacková · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Růžena Sedláková · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9

Email triage · German (150 documents)

GPT-6 Sol is more accurate by 3.1 pp. Gemma 4 31B is 18.4× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	95.1%	98.2%
Error-free documents	62%	84%
Per 1,000 documents	$0.15	$2.73
To fix / 1,000	380	160
Speed (median)	3.6 sec	2.9 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Category	100.0%	100.0%
Priority	78.0%	97.3%
Sentiment	78.0%	86.7%
Customer name	100.0%	100.0%
Order number	100.0%	100.0%
Invoice number	100.0%	100.0%
Amount	100.0%	100.0%
Currency	100.0%	100.0%
Deadline	100.0%	100.0%

Documents where they differ

Alexej Sauer · Email · Gemma 4 31B 7/9, GPT-6 Sol 9/9
Ibrahim Walter · Email · Gemma 4 31B 7/9, GPT-6 Sol 9/9
Sylwia Gerlach · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Benedikt Speer · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Cristina Krause · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Cynthia Hahn · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Roselinde Heinz · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9
Linda Bähr · Email · Gemma 4 31B 8/9, GPT-6 Sol 9/9

Personal data detection · English (US) (150 documents)

Gemma 4 31B is more accurate by 0.2 pp. Gemma 4 31B is 24.8× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	99.9%	99.7%
Error-free documents	99%	98%
Per 1,000 documents	$0.08	$2.09
To fix / 1,000	7	20
Speed (median)	2.4 sec	2.6 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Names	100.0%	100.0%
Email addresses	100.0%	100.0%
Phone numbers	100.0%	100.0%
Addresses	99.3%	98.0%
Dates of birth	100.0%	100.0%
National IDs	100.0%	100.0%
Bank accounts	100.0%	100.0%

Documents where they differ

Both models got the same number of fields right on every example document.

Personal data detection · Czech (150 documents)

Gemma 4 31B is more accurate by 0.1 pp. Gemma 4 31B is 18.7× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	99.8%	99.7%
Error-free documents	99%	98%
Per 1,000 documents	$0.12	$2.22
To fix / 1,000	13	20
Speed (median)	2.9 sec	2.5 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Names	100.0%	100.0%
Email addresses	100.0%	100.0%
Phone numbers	100.0%	100.0%
Addresses	98.7%	98.0%
Dates of birth	100.0%	100.0%
National IDs	100.0%	100.0%
Bank accounts	100.0%	100.0%

Documents where they differ

Both models got the same number of fields right on every example document.

Personal data detection · German (150 documents)

Gemma 4 31B is more accurate by 0.2 pp. Gemma 4 31B is 20.9× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	99.9%	99.7%
Error-free documents	99%	98%
Per 1,000 documents	$0.10	$2.13
To fix / 1,000	7	20
Speed (median)	1.9 sec	2.6 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Names	100.0%	100.0%
Email addresses	100.0%	100.0%
Phone numbers	100.0%	100.0%
Addresses	99.3%	98.0%
Dates of birth	100.0%	100.0%
National IDs	100.0%	100.0%
Bank accounts	100.0%	100.0%

Documents where they differ

Gunnar Zorbach · Text · Gemma 4 31B 6/7, GPT-6 Sol 7/7

Contract clauses · English (US) (150 documents)

GPT-6 Sol is more accurate by 0.4 pp. Gemma 4 31B is 18.5× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	99.6%	100.0%
Error-free documents	97%	100%
Per 1,000 documents	$0.18	$3.33
To fix / 1,000	27	0
Speed (median)	2.5 sec	2.5 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Provider	98.0%	100.0%
Provider ID	100.0%	100.0%
Client	98.0%	100.0%
Client ID	100.0%	100.0%
Contract type	100.0%	100.0%
Price (net)	100.0%	100.0%
Currency	100.0%	100.0%
Billing period	100.0%	100.0%
Effective date	99.3%	100.0%
End date	100.0%	100.0%
Notice period (months)	100.0%	100.0%
Late-payment penalty (% per day)	100.0%	100.0%
Governing law	100.0%	100.0%

Documents where they differ

Both models got the same number of fields right on every example document.

Contract clauses · Czech (150 documents)

Both are equally accurate. Gemma 4 31B is 18.2× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	100.0%	100.0%
Error-free documents	99%	100%
Per 1,000 documents	$0.25	$4.62
To fix / 1,000	7	0
Speed (median)	6.0 sec	2.0 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Provider	100.0%	100.0%
Provider ID	100.0%	100.0%
Client	100.0%	100.0%
Client ID	100.0%	100.0%
Contract type	100.0%	100.0%
Price (net)	100.0%	100.0%
Currency	100.0%	100.0%
Billing period	100.0%	100.0%
Effective date	99.3%	100.0%
End date	100.0%	100.0%
Notice period (months)	100.0%	100.0%
Late-payment penalty (% per day)	100.0%	100.0%
Governing law	100.0%	100.0%

Documents where they differ

Elektrotrade s.r.o. · Konečný a Dostál spol. s r.o. · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13

Contract clauses · German (150 documents)

GPT-6 Sol is more accurate by 5.6 pp. Gemma 4 31B is 19.8× cheaper.

Metric	Gemma 4 31B	GPT-6 Sol
Correct fields	94.4%	100.0%
Error-free documents	27%	100%
Per 1,000 documents	$0.22	$4.36
To fix / 1,000	727	0
Speed (median)	2.6 sec	2.9 sec
Valid answers	100.0%	100.0%

Accuracy per field

Field	Gemma 4 31B	GPT-6 Sol
Provider	100.0%	100.0%
Provider ID	100.0%	100.0%
Client	100.0%	100.0%
Client ID	100.0%	100.0%
Contract type	100.0%	100.0%
Price (net)	100.0%	100.0%
Currency	100.0%	100.0%
Billing period	100.0%	100.0%
Effective date	100.0%	100.0%
End date	100.0%	100.0%
Notice period (months)	100.0%	100.0%
Late-payment penalty (% per day)	27.3%	100.0%
Governing law	100.0%	100.0%

Documents where they differ

Hahn GmbH · Bauer & Klein GmbH · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Wesermetall e.K. · Weserform GmbH · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Haustechnik Braun UG (haftungsbeschränkt) · Nordbau GmbH · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Hansawerk GmbH · Schreinerei Weber AG · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Maintec GmbH · Schreinerei Wagner GmbH · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Winkler & Beck e.K. · Meyer & Weber e.K. · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Kraus Consulting GmbH · Lange Consulting GmbH & Co. KG · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13
Hansametall GmbH & Co. KG · Roth & Schröder UG (haftungsbeschränkt) · Contract text · Gemma 4 31B 12/13, GPT-6 Sol 13/13