There are thousands of evals to judge the best AI. But the best one is: What are people paying for? Using OpenRouter usage data, we can calculate the API revenue of the top models:
Estimated revenue per model using OpenRouter pricing. Recent weeks use a complete per-model feed that reports every model separately with prompt and completion tokens split out, so revenue is computed directly. Earlier weeks come from a feed that broke out only the top nine models per week and pooled the rest into an “Others” bucket worth roughly a third of all tokens; there, revenue assumes 1.5% completion tokens and the pooled tokens are valued at the blended rate of the named models. Completed promotions can be back-adjusted using post-promo baselines when configured. Dashed vertical rules mark weeks where a large model dropped out of the old breakdown, which reads as a drop but is a reporting change; the solid orange rule marks where the complete feed takes over.
And from this, we can get a rough estimate of the relative API usage dominance of the different labs:
Estimated revenue per lab using OpenRouter pricing. Recent weeks use a complete per-model feed that reports every model separately with prompt and completion tokens split out, so revenue is computed directly. Earlier weeks come from a feed that broke out only the top nine models per week and pooled the rest into an “Others” bucket worth roughly a third of all tokens; there, revenue assumes 1.5% completion tokens and the pooled tokens are valued at the blended rate of the named models. Completed promotions can be back-adjusted using post-promo baselines when configured. Dashed vertical rules mark weeks where a large model dropped out of the old breakdown, which reads as a drop but is a reporting change; the solid orange rule marks where the complete feed takes over.
Normalizing to 100% makes it easier to see how each lab's share of the pie is changing over time:
Each lab's estimated revenue as a percentage of the weekly total across labs the source names individually. The untracked “Others” bucket is left out of this denominator so the comparison stays readable; it is included in the absolute charts above. Dashed rules mark weeks where a large model dropped out of the breakdown, which lifts every remaining lab's share mechanically.
We also need to look into the future to see what kind of compute each lab will have access to. Using satellite imagery and public records, Epoch AI tracks the frontier data centers being built by each lab. Here's how much compute each lab will have access to over the next few years: