Rankings methodology

A usage leaderboard is only as honest as its outliers. In early 2026, a free promotional model reached #1 on a major aggregator’s rankings on the strength of a single customer routing trillions of tokens — technically accurate, practically meaningless. Our rankings are built so that can’t happen here.

What the rankings measure

Tokens routed through OpenKey per model, aggregated weekly. That’s a demand signal from paying workloads — not a quality benchmark. For capability comparisons, model pages carry third-party benchmark data (Design Arena, Artificial Analysis) clearly attributed.

The filters

1. Whale filtering. A single account’s traffic is capped at 5% of any model’s counted volume. One customer, however large, cannot move a ranking by more than one position band.
2. Promo-price flags. Models priced at zero or steeply discounted by their provider are flagged on the leaderboard. Free traffic is ranked separately from paid traffic — free usage tells you what’s popular to experiment with, paid usage tells you what people trust with money.
3. No self-dealing. Traffic generated by OpenKey’s own systems (health checks, playground demos, verification pipelines) is excluded.
4. Alias folding. Rolling aliases (*-latest) count toward their underlying model, not as separate entries.

What we publish

Each ranking shows: counted token volume, the number of distinct accounts behind it, week-over-week change, and any active flags. If a methodology change alters historical rankings, the change is dated in the changelog and old snapshots stay available.

What we don’t do

We don’t rank by revenue (it privileges expensive models), we don’t accept payment for placement, and we don’t editorially pin models. If a lab believes its numbers are wrong, [email protected] — disputes and resolutions are logged publicly.