Which model answers your customers best?

Upload the questions your users really ask, pick two models, and get a report that tells you which one is better, which one is cheaper, and how sure the result is.

Start a test Request an access code

B winssame verdict in both orders

Question 7 of 20, refundsCan I get a refund 20 days after purchase?

Answer ARefunds may be possible. Please contact support for details.

Answer BYes. Refunds are accepted within 30 days of purchase with the original receipt.

Accuracy

Completeness

Cost / answer

The report answers the questions a model decision actually needs

Best quality: Scored against your rubric and your reference answers, never guessed when there is no reference.
Cheapest that is good enough: Real cost per answer from the provider, projected to your monthly volume.
Best value: The cheapest model within 5% of the leader, if there is one.
How sure we are: Sample size, judge consistency and errors are reported. Too little evidence means no winner is declared.

How it works

Upload your questions
A simple CSV with the questions your customers really ask. Add the correct answer or a policy text when you have one.
Pick two models and a judge
Choose the models you are considering. A model from another company judges the answers without knowing who wrote them.
Read the verdict
You get the best model, the cheapest one that is good enough, the best value, and how sure we are.

What is inside the report

Three clear decisions, each with the evidence behind it
Quality per model and per criterion: accuracy, completeness, tone and format
Head-to-head results, every pair judged twice with the answers swapped
Real cost per answer and a projection for your monthly volume
The lowest-scored answers with the judge’s reasoning
A warning when the evidence is thin, instead of a made-up winner

Built so the result can be trusted

Blind judging: The judge never sees which model wrote an answer and treats the text it grades as data, not as instructions.
Order swapped: Every head-to-head comparison runs twice with the answers swapped. Disagreement is reported as uncertainty, not as a tie.
Honest about evidence: Without a reference answer or context, accuracy is marked as not assessed instead of guessed.
Your data: Questions and reports are deleted after 30 days. Do not upload personal or confidential data.

Questions

What is an access code?

Each code runs one complete test, paid by us, with up to 20 questions and 2 models. Write to us to get one.

How long does a test take?

It depends on the number of questions and the models. You can close the page; the report is here when you come back.

Which models can I test?

Most models available on OpenRouter. Start typing in the model field to see the list with prices.

In which language is the report?

You choose, English or Turkish, independent of the language of this page.

What if a test fails?

If the failure is on our side, your code is not used and you can run it again.

Ready to compare your models?

Bring a CSV with your real questions and get a verdict you can act on.