Who's going to cancel?
We draw customers from a real telecom dataset. Each model gets the profile — without knowing what happened — and estimates the chance of cancelling. Then we compare with reality. The race is between AI models; a logistic regression trained on this data joins as the classic-method reference.
Heads-up: the tasks themselves run in Portuguese (Brazilian news feed, Portuguese words, same prompts for every model). That keeps every run comparable — and the leaderboard in one piece.
Start here: let's draw the customers
This test runs on a real telecom dataset. We draw customers at random — you get to see who's in the sample before any model weighs in.
Each customer becomes a separate request per model, run in sequence — the fairest way to measure. Models see the profile (tenure, contract, services, charges), never the ID or the Churn column. If a model takes longer than 20 seconds, we pull it from the race and the report moves on with whatever it already delivered.
Now just kick off the race
Telco Customer Churn, from IBM — one of the classic machine learning datasets, with 7,043 customers from a fictional California telecom company. Publicly available on Kaggle. Here we only use the Churn column as the answer key: it's never sent to the models — click any model's logo to see the exact request. The reference logistic regression is retrained on this same dataset every session; the LLMs never see any data from the population. Click the Σ to see what it learned.