Publication

Model summary · Edition 1.1

Gemini 3.7 FlashHigh

These results come from one run: 10 fixed subjects, 3 rounds each, a 40-question limit, and reasoning setting high. They do not predict results on other subjects or settings.

This page always shows the model’s newest ranked run.

Every question, answer, guess, and source.

Result

Rank
Rank 7of 20 models
Question score
16.57 Lower is better. 95% confidence interval 14.35 to 18.79.
Rounds solved
28 of 30 0 format errors
Guesser cost per round
$0.0525 Model time 52s.
See how Gemini 3.7 Flash (high) played. All 30 rounds, with every question, answer, guess, and source.

Observations

What stood out in this run.

Counts and ranks from this run only. A subject rank compares this model’s average on that subject with every other model in this edition. Each item links to its rounds.

Best ranks vs other models

Lowest ranks vs other models

Single rounds

  • Fewest questions in a solved round: 8, on Achilles. round 1
  • Most questions in a solved round: 32, on Door handle. round 3
  • Largest spread between rounds: Bike pump, from 23 to 41 questions. round 1, round 2, round 3
  • 4 wrong guesses in 30 rounds, up to 2 in one round on Bike pump. round 3

Failed rounds and format errors

  • Not solved: Bike pump. Question limit reached after 40 questions; scored as 41. round 3
  • Not solved: Door handle. Question limit reached after 40 questions; scored as 41. round 2

Subjects

Results by subject.

Failed rounds count as 41 questions. Rank compares this subject's average with the other models in edition 1.1. Select a subject or a round to see the game.

SubjectAverage questionsSubject rankSolvedRounds
Albert Einstein12.0018 of 203 of 3
Albert Schweitzer18.003 of 203 of 3
Garfield11.677 of 203 of 3
Achilles8.339 of 203 of 3
Genghis Khan12.009 of 203 of 3
Bike pump29.677 of 202 of 3
Spider web12.672 of 203 of 3
Eyebrow18.3312 of 203 of 3
Moon11.0016 of 203 of 3
Door handle32.0014 of 202 of 3

Runs

Runs of this model.

Each edition ranks only the newest complete run of a model. Earlier runs are not ranked and their details are not published.