Why the model ranked #2 might really be #1: reading score intervals
Leaderboards love a clean ordering: first, second, third. But each score is an estimate built from a limited number of benchmark results, and estimates come with uncertainty.
What the interval means
ModelBench shows a 90% interval next to every overall score, for example 82.2 (77–88). Roughly: given the evidence available, the model's true score is very likely somewhere in that range.
Overlap means "not distinguishable"
Take two models:
| Model | Score | 90% interval |
|---|---|---|
| A | 86.5 | 82–91 |
| B | 84.1 | 80–88 |
A is listed first, but the ranges overlap heavily. The honest reading is: A and B are about equally capable on this evidence. The FAQ on our home page calls this out automatically when the top two overlap.
Why intervals differ in width
A model with results on many benchmarks has a narrow interval; a new model with only a handful of reported results has a wide one. That is also why some models are marked Provisional: some of the evidence behind the score is still being sourced.
Practical advice
- Use the score to build a shortlist, not to pick a single winner.
- Within the shortlist, decide on what you can measure precisely: price, context window, speed, and how the model does on your tasks.
- Check the category tabs (coding, math, agentic…). A model that is third overall can be first at the thing you need.