Insights
9.6 trillion tokens a day: ranking models by what actually runs
Ask ten people which AI models "everyone uses" and you get ten guesses. TKX's heat board no longer guesses: it ranks models by measured token volume — real traffic, read daily from the largest public routing marketplace. On July 16, 2026, that meant 9.62 trillion tokens in a single day.
The top of the board, measured
The ten busiest models by routed token volume (July 16, 2026):
| # | Model | Tokens / day | Share |
|---|---|---|---|
| 1 | hy3 | 1,745 B | 18.1% |
| 2 | mimo-v2.5 | 1,387 B | 14.4% |
| 3 | deepseek-v4-flash | 777 B | 8.1% |
| 4 | glm-5.2 | 624 B | 6.5% |
| 5 | minimax-m3 | 564 B | 5.9% |
| 6 | nemotron-3-ultra-550b | 441 B | 4.6% |
| 7 | deepseek-v4-pro | 392 B | 4.1% |
| 8 | claude-4.7-opus | 390 B | 4.1% |
| 9 | claude-4.8-opus | 332 B | 3.5% |
| 10 | claude-4.6-sonnet | 171 B | 1.8% |
Live data, refreshed daily: /api/v1/usage.json.
What the data shows
- Open-weight models dominate routed traffic. The top five spots — more than half of all routed volume — are open-weight or aggressively priced models. On a price-transparent marketplace, cheap-and-good wins the bulk workloads.
- Frontier closed models still buy their place. Three Claude variants hold ~9% of routed volume combined at premium prices — volume that persists at a price premium is the clearest willingness-to-pay signal in the dataset.
- Usage concentrates hard. The top 10 models carry roughly 70% of all routed tokens; the long tail of 300+ listed models shares the rest. Catalogue size and actual usage are very different things — which is exactly why the heat board now ranks by the latter.
The honest scope note
This is real usage on one marketplace, not global market share. OpenRouter's traffic skews price-sensitive and multi-model; most first-party API traffic (e.g. direct enterprise use of a vendor's own API) never touches a router and is invisible here. So closed flagship models are under-represented in these numbers. TKX labels the metric accordingly — a measured slice, honestly scoped, beats a guessed total.
Why rank by usage at all?
Because the alternative is worse. Popularity proxies — search volume, social buzz, benchmark scores — are all gameable and none of them clear at a price. A routed token is a paid, delivered unit of work. It is the same discipline that runs every TKX board: measure, tag the evidence, and never accept self-reported numbers.
← All insights