AI Token Usage Tracker
Daily token consumption across the entire OpenRouter platform: weekly totals and growth, plus shifting market share between AI providers. Token volume is the most direct high-frequency proxy for real AI demand.
Latest day
30.61 T
2026-10-22
7-day average
24.87 T
vs previous 7 days -2.3%
Last full week
164.0 T
2026-W42 · WoW -4.5%
WoW (4-week smoothed)
+5.7%
Week over week, smoothed
Year over year
+3432%
vs the same week a year ago
Weekly token consumption
Aggregated by calendar week (Monday start). Top 50 models plus the other long-tail aggregate row equals the true platform total.
An unfinished final week understates the total and is excluded from growth rates.
Growth rate
This chart answers whether AI demand is accelerating or decelerating. Teal is the 4-week-smoothed week-over-week rate (short-term momentum); brown is year over year (long-term slope). The YoY figures are dominated by base effects — read the direction, not the absolute value.
Week over week (4-week smoothed, left axis)Year over year (vs same week last year, right axis)
Market share by provider
Token share aggregated by provider. Only the highest-volume providers get their own colour; everything else is folded into a grey "other providers" bucket. Share migration often carries more signal than total growth.
Shares are normalised to the platform total for that week; grey is the combined remainder of unhighlighted providers.
Provider share: latest day vs one year ago
Side-by-side ranking at two points in time. Note: a falling share often just means traffic moved to first-party APIs — it does not necessarily mean demand disappeared.
2026-10-22 · 30.61 T that day
| Provider | tokens | share |
|---|
| DeepSeek | 9.43 T | 30.8% |
| MiniMax | 3.52 T | 11.5% |
| 智谱 AI | 2.98 T | 9.7% |
| Anthropic | 2.76 T | 9.0% |
| OpenAI | 1.96 T | 6.4% |
| 腾讯 | 1.92 T | 6.3% |
| 其他长尾 | 1.83 T | 6.0% |
| NVIDIA | 1.24 T | 4.0% |
| 小米 | 1.16 T | 3.8% |
| Google | 0.80 T | 2.6% |
| 其余厂商合计 | 3.01 T | 9.8% |
2025-10-22 · 0.7 T that day
| Provider | tokens | share |
|---|
| xAI | 0.2 T | 29.6% |
| Google | 0.1 T | 18.2% |
| Anthropic | 0.1 T | 18.1% |
| OpenAI | 0.1 T | 10.3% |
| 其他长尾 | 0.1 T | 7.1% |
| DeepSeek | 0.0 T | 5.7% |
| 阿里 Qwen | 0.0 T | 3.1% |
| 智谱 AI | 0.0 T | 2.7% |
| TNG Tech | 0.0 T | 2.0% |
| Mistral AI | 0.0 T | 1.1% |
| 其余厂商合计 | 0.0 T | 2.1% |
Providers beyond the top 10 are combined into "all other providers"; "long tail" is the aggregate row shipped by the data source itself.
What token consumption is, and why track it
Token consumption measures how much AI models are actually being called — a number far harder to inflate than funding rounds or launch keynotes. TokenDealHub pulls the official OpenRouter public dataset every day and breaks platform-wide token usage down by calendar week and by provider. The platform currently processes roughly 26 trillion tokens per day, up from about 0.7 trillion a year ago — roughly 37x growth. It is the most direct high-frequency proxy for real AI demand and competitive dynamics.
Frequently asked questions
What exactly does this page measure?
It comes from the OpenRouter official public dataset and counts every token processed through the OpenRouter API (both prompt and completion). The dataset returns the top 50 models each day plus one other aggregate row covering rank 51 and beyond, so the sum is the true platform total with no long tail lost. Requests marked private are excluded.
Why weekly instead of daily?
Day-of-week effects dominate: weekends typically run below weekdays, so a daily series is a sawtooth that hides turning points. Aggregating by calendar week gives a stable trend; use the week-over-week and year-over-year curves when you need finer resolution.
How should I read provider market share?
Read the direction, not the absolute value. A rising share means developers and applications are migrating toward that provider. A falling share should first be checked for channel effects such as a move to first-party APIs before you conclude demand changed. Model version swaps and launches also move single-model share, which is why provider-level aggregation is more reliable than model-level.
Can token consumption be used as an AI investment indicator?
Yes, but only as one demand-side leading indicator, and only alongside price. Token volume reflects call volume, price reflects unit value, and the product approximates revenue. It is also a single-platform sample (roughly 1% of global usage) skewed toward developers and individuals — enterprise contracts do not flow through here. Cross-check it against cloud capex, inference price curves, and the token figures providers disclose themselves.
How to read this
- This covers OpenRouter only — roughly 1% of global token consumption, skewed toward developers and individual users. It reveals competitive dynamics, not the global total.
- Tokens are not revenue. Inference prices are falling fast, so 10x the tokens may mean only 2x the revenue. Reading volume alone overstates industry health.
- Year-over-year growth is dominated by base effects: the early baseline is tiny, so YoY looks absurdly high. Judge the trend by whether its downward slope is flattening or turning, not by the absolute value.
- A falling provider share does not mean disappearing demand. A common cause is traffic moving to first-party APIs or enterprise contracts, leaving the OpenRouter sample pool.
- Each provider uses its own tokenizer, so the same text can differ by 2x in token count. Summing across providers inherently carries noise.
Source: OpenRouter official public dataset (Data API, CC BY 4.0). Please credit OpenRouter and TokenDealHub when republishing.
