SnapSheet

Benchmark

Finding the best AI
for product descriptions.

305 head-to-head votes across 14 real products. Voters chose between two descriptions without knowing which model wrote each.

Goal

Our goal was to build an API that lets you easily implement an image-to-description pipeline for a second-hand marketplace/exchange, specifically for the "TBZ Tauschhütte".

305

Human votes

14

Products tested

8

Models compared

ELO ranking

Average ELO across all products. Starting ELO is 1400.

1 Gemini Flash
1441
single
2 Gemini→Claude
1426
pipeline
3 GPT-5 Nano
1417
single
4 Claude Sonnet 4
1416
single
5 GPT-4o Mini→4o
1415
pipeline
6 Qwen 3.5
1393
single
7 Mistral Small
1363
single
8 GPT-4o
1328
single

Win rate

Head-to-head matchup win percentage across all products.

Gemini Flash
88.5%

46W / 6L

Gemini→Claude
71.9%

41W / 16L

GPT-5 Nano
65.5%

36W / 19L

Claude Sonnet 4
63.5%

33W / 19L

GPT-4o Mini→4o
66.7%

30W / 15L

Qwen 3.5
45.5%

30W / 36L

Mistral Small
23.9%

16W / 51L

GPT-4o
1.4%

1W / 71L

Votes won

Total times each model was preferred by voters.

59
48
53
50
41
30
16
5
Gemini Flash
Gemini→Claude
GPT-5 Nano
Claude Sonnet 4
GPT-4o Mini→4o
Qwen 3.5
Mistral Small
GPT-4o

Pipelines

Splitting image analysis and copywriting across two models.

Our idea with pipelines was that we could leverage the image parsing capabilities of one model with the writing competence of another. Our testing revealed that even though it is possible and produces great results, it doesn't beat out just using a model directly. We speculated that it's because a lot of context is lost when the data is handed over from one model to another.

Performance by product

ELO per model per product. Colors show relative rank within each row — green beats red.

ProductGemini FlashGemini→ClaudeGPT-5 NanoClaude Sonnet 4GPT-4o Mini→4oQwen 3.5Mistral SmallGPT-4o
Rollschuhe
1498
1405
1453
1404
1426
1328
1357
1329
Cars Carrara Bahn
1497
1385
1383
1416
1374
1401
1416
1328
Radon Trekkingrad Solution Sport 3
1484
1403
1381
1430
1414
1380
1357
1351
Schreibtisch
1431
1474
1429
1377
1427
1370
1326
1366
Kinderwagen
1460
1431
1428
1416
1416
1382
1343
1324
Petroleum Lampe
1457
1416
1417
1404
1442
1394
1341
1329
Keramik-Wanduhr
1430
1432
1455
1356
1368
1445
1383
1331
Zangen Set
1419
1440
1443
1384
1455
1408
1357
1294
Waschmaschine
1413
1428
1432
1384
1431
1416
1382
1314
Roter Sessel
1443
1399
1385
1444
1416
1415
1356
1342
Monopoly Bremen
1399
1417
1448
1447
1412
1383
1395
1299
Bosch Bohrmaschine
1413
1442
1456
1458
1386
1385
1322
1338
Bowleset mit Tassen
1404
1482
1348
1458
1415
1383
1366
1344
Pegasus Fahrrad
1429
1409
1384
1446
1433
1413
1385
1301

Limitations

Due to our limited budget, we were only able to test and select a limited number of models. This means that there might be a better model out there, but using it would not be economically viable.

Furthermore, we had to carefully pick out what items we tested, we based our selection on what would usually be found in the "Tauschhütte" and on second-hand marketplaces like "kleinanzeigen.de". In combination with the fact that we didn't test specific model-to-product combinations, it is possible that specific products would perform better for product types than our overall best config (Gemini Flash).

Methodology

ELO rating

All models start at 1400. Each vote adjusts both scores with the standard ELO formula (K=32).

Blind voting

Voters see two descriptions without knowing which model wrote each. They pick the one they prefer.

Coverage

Every possible pair of descriptions for each product was shown as a matchup.

Snapshot: 2026-06-03 · 305 votes · 14 products