Test AI models v2.0 is hereSee what's new
Test AI models
Blog / Case study

14 AI models on 5 freight documents: what we found

Zonik AI, a logistics product we co-founded, compared 14 AI models on 5 documents from 3 broker templates in Test AI models. What it found, and what it does not show.

Marko MilojkovicFounder, Test AI models
3 min read

Key takeaways

  • Zonik AI, a logistics product we co-founded, compared 14 AI models on real US broker rate confirmations in Test AI models: 3 scanned and 2 digital documents, from 3 broker templates.
  • On the scanned documents, the two winning models each cost 83-89% less than gpt-4o, at list prices on 2026-09-16.
  • The strongest model on scanned documents failed both digital documents, so no single model won both kinds.
  • The test also found 7 reproducible extraction errors, and they were fixed in the production prompt.

In September 2026, 14 AI models read the same real US broker rate confirmations in Test AI models. The two winners on scanned documents each cost 83-89% less than gpt-4o, at list prices on 2026-09-16.

The test ran for Zonik AI, a logistics product we co-founded. The sample was small: 3 scanned and 2 digital documents, from 3 broker templates.

The full write-up, with every result, is on the Eterna blog: We tested 14 AI models on freight rate confirmations (opens in a new tab). This post covers what the test found, how the costs were worked out, and what it does not show.

The test was 14 models on 5 documents

  • 14 AI models. The model names and prices are as they were in September 2026.
  • 3 scanned documents and 2 digital documents. The results for each kind are reported separately.
  • 3 broker templates. All were real US broker rate confirmations.

On scanned documents, two cheaper models won

On the 3 scanned documents, Gemini 3.1 Flash Lite Preview and Grok 4 Fast were the winners. Each cost 83-89% less than gpt-4o, at list prices on 2026-09-16.

At those list prices, running both scanned-document models still cost less per document than one gpt-4o call in this test.

No single model won both kinds of document

On the 2 digital documents, Grok 4 Fast and DeepSeek V3.2 were the winners.

The strongest model on the scanned documents failed both digital documents. On this sample, the right model depended on the kind of document.

The test also fixed the prompt

The test found 7 reproducible extraction errors. They were fixed in the production prompt.

How the costs were worked out

Costs are the tokens each model used in the test, priced at each provider's list prices on 2026-09-16. They are list prices on that day, not what anyone paid in production.

What Zonik AI runs now

Since 2026-09-18, Zonik AI runs two models on the rate confirmations its first users upload: Gemini 3.1 Flash Lite Preview and Grok 4 Fast on scanned documents, and Grok 4 Fast and DeepSeek V3.2 on digital ones. The dispatcher reviews and confirms before the load is saved.

Zonik AI's two-model check compares the two answers field by field: dates, times, rate, weight, and miles must match exactly, the broker name is compared after normalizing it, and notes and labels are not compared. A field where the two models disagree is flagged for the dispatcher.

The limits

  • It is 5 documents from 3 broker templates. That is a small sample, not a ranking of AI models.
  • Costs are list prices on 2026-09-16. Prices change, and the model names will too.
  • It is not a production saving. On the digital documents no saving is claimed, because gpt-4o-mini was not in the test.

Run a test like this on your own documents

Test AI models 2.0 has two features for a test like this. The quality check marks each answer against a rule you write, and Try cheaper models runs the same prompt on cheaper models you have not tried, judged by the same rule. Our guide shows how to write that rule: How to write a pass rule for the quality check.

Find the cheaper model for your own prompt
  • 10k free credits
  • No card
Run your own test

Marko

Questions

Which models won?

On the scanned documents, Gemini 3.1 Flash Lite Preview and Grok 4 Fast. On the digital documents, Grok 4 Fast and DeepSeek V3.2. That is on 5 documents from 3 broker templates.

Is Zonik AI an independent customer?

No. Zonik AI is a logistics product we co-founded, and Test AI models is our own product.

Will the same models win on your documents?

Not necessarily. This was one document type and a small sample. Run your own documents and your own prompt before you switch models.

Marko MilojkovicFounder, Test AI models

Runs Eterna Creative and built Test AI models after repeated multi-provider LLM integration work for client products.

About the team