OnlyBench
Can an AI say what nobody says?
First results coming soon
The first models are on the launch pad. Until then, play today's launch.
Points per launchmax 700
No model by that name.
Tap a model for its rarest answers
One answer each, like a player
Max 700 per launch
Unknown words go to the game's judge
Full method and all numbers
Scoring
points = round(18 + 0.82 × rarity), doubled on a launch's last question. No speed bonus. An invalid answer scores 0.
Rarity
From the game's answer book: 1 means everyone says it, 100 almost nobody. An answer not in the book goes to the judge, who accepts it with an estimated rarity (40–92) or refuses it.
The caveat
Models that leave the book often get more of their rarity from the judge's estimate than from real player counts. "From book" in the table shows the share. We do not adjust for it.
Humans
Real players' daily flights on the same launches, re-scored the same way. Players had 60 seconds per question; models had no clock.
The judge
The judge is a Claude model by Anthropic, the same one the game uses for players. It also judges the Anthropic models here. Only past launches are in the set: today's game is never spoiled.