aimodelranksStart
HOW THE RANKINGS ARE MADE
Every AI leaderboard is a choice of test.
Some rank models on fixed sets of questions, others on thousands of blind head-to-head votes. Researchers who study them say the method, the margin and the fine print matter as much as the number one spot.
MIND THE GAPbetween first and second
Check a rankingKeep in mind
THE SHORT ANSWER
Most AI rankings come from one of two methods.
Each has strengths and known weak spots.
TEST SETS
Many benchmarks score models on a set list of questions. If similar questions leak into training data, scores can look better than the skill behind them. On a fresh maths set, researchers saw drops of up to 8%.
Check a rankingHEAD TO HEAD
Arena-style leaderboards show people two anonymous answers and ask which is better. One such arena had over 240,000 votes in more than 100 languages by January 2024.
See the numbersCLOSE CALLS
Stanford's AI Index found the gap between the first and tenth-ranked models fell from 11.9% to 5.4% in a year, with the top two just 0.7% apart.
How to read oneA QUICK RANKING CHECK
Pick the kind of ranking in front of you to see what research says about it. The bar shows how much care to take before reading much into the order.
Based on published research and the UK AI Security Institute. General information about how rankings are made, not a verdict on any model.
TWO NUMBERS WORTH KNOWING
One figure on how tight the top of the table is,
one on what happens when the questions are new.
0.7%
Gap between first and second
Stanford's 2025 AI Index says the score difference between the top and 10th-ranked models fell from 11.9% to 5.4% in a year, and the top two were separated by just 0.7%.
THE TOP TWO · AI INDEX 2025
up to 8%
Accuracy drop on a new maths set
Researchers commissioned new grade-school maths problems to mirror a popular benchmark. Some models scored up to 8% lower, though many frontier models showed minimal signs of overfitting.
FRESH QUESTIONS · GSM1K
The GSM1k authors also found that models more likely to reproduce examples from the original test tended to show bigger gaps, suggesting some had partly memorised it.
FOUR HABITS TO TRY
None of this needs technical knowledge. It is about asking what was measured, how big the gaps are and who had a chance to tune for the test.
Step 1
Maths, coding, writing or general chat? A model that leads one benchmark may not lead another.
Step 2
If the scores or confidence ranges overlap, treat the models as roughly level.
Step 3
Tests written after a model was trained, or kept private, are harder to memorise.
Step 4
Public rankings average many people's questions. A few of your own real tasks tell you more about fit.
WHY RANKINGS DRIFT
Once a leaderboard becomes famous, developers have reasons to tune for it. The 2025 study The Leaderboard Illusion found that some providers tested many private variants before release, and that closed models were sampled in more battles than open ones.
The authors estimated that even limited extra arena data could produce relative gains of up to 112% on the arena's own question mix, and warned this rewards fitting the arena rather than general quality.
Questions
A fixed set of tasks or questions used to score models in the same way. Stanford's HELM, for example, scored 30 models on the same core scenarios under standardised conditions so they could be compared fairly.
A person asks a question, sees answers from two anonymous models and picks the better one. The votes feed a statistical model, the Bradley-Terry model, which turns many pairwise results into scores with confidence intervals.
It is when questions from a test, or very similar ones, leak into a model's training data. The model can then score well partly from memory rather than skill, which is why researchers build fresh test sets.
Not necessarily. Rankings average particular tests or many people's questions. The AI Index notes the top models are now very close, so trying a few of your own tasks often tells you more.
No. The UK AI Security Institute tests systems for capabilities and risks, and says these evaluations give a snapshot of a system rather than a certificate that it is safe.
Destination
Updated
ABOUT THIS NAME
aimodelranks.com is part of the DoodleSale portfolio. It isn't in an auction right now, but you can ask about leasing it.
