The Arena
One prompt, two models, side by side and unnamed. You pick. Then both names and their receipts.
Ask one question
The router picks two models from the free tier or the live catalogue. You never see which is which until you vote.
Sign in to askTwo models answer blind; your vote is yours.
How it works
- Two answers, no names. The same prompt runs on two models and you see only the words, side by side.
- You vote once. Left, right, a tie, or both bad. One vote per person per round, and a second is refused.
- Then the reveal. Both names, the offer that answered, the price of the run, its speed and the router's reason.
- It counts. Your vote is an OpenVibe.Bench human_pairwise judgement, filed under the kind of work your prompt asked for: code, writing, chat, reasoning, translation or analysis.
The rankings put those ratings beside the price per million tokens, the measured speed and the uptime, so “worth it” is visible. One vote is one sample: the board marks a model provisional until it has enough games.
It is public. An arena prompt and both answers are recorded in Bench's open arena dataset, without your name, so anyone can check a vote. Do not paste anything private.
10 models in the catalogue; two are chosen for each prompt.