OpenRouter · tokens processed per week · Feb 2 – Sep 22, 2026
Is OpenRouter's token usage growing exponentially?
Yes, and slightly faster than exponential. Weekly tokens grew 13.1× in 32 weeks. A constant 7.6% weekly growth rate (an exponential curve) fits with R² 0.96, against 0.83 for a straight line. The rate has also gone up: 6.3%/week from February to mid-May, then 8.9%/week since.
This page recreates mk7s/exporouter with Holochart, then looks at the same growth from other angles. Every chart below is a Holochart example that runs live in your browser and doubles as a visual regression test. The original draws its main chart as an orbitable 3D scene; Holochart's 3D scenes arrive in M6, so here it is 2D.
Weekly tokens
Hover a bar for the week's detail and top models. Click a fit in the legend to hide it, and switch the y axis to a log scale, where an exponential is a straight line.
If the trend holds, the fit points to 157T/week by Oct 12. OpenRouter's own forecast for the current week (two days in) is 140T, compared with 125T from the fit.
Is the growth rate holding up?
Rolling 4-week compound growth, per week. A pure exponential would hover around the dashed line; bars trending above it mean faster than exponential.
How this was measured
- Source: the Top Models weekly chart on OpenRouter's rankings page, summed across every model (the top 9 plus "Others"), so each bar is total text-model tokens (prompt plus completion) for a Monday-start week.
- The fits use the 33 complete weeks, Feb 2 to Sep 14. The week of Sep 21 only covers 2 days, so it's drawn but left out of the fits.
- Exponential fit: least squares on
ln(tokens)against the week index, givingtokens ≈ A·e^kt. The linear and quadratic fits use raw tokens. R² is measured in raw token space for all three, so they compare fairly. - AIC penalises the quadratic's extra parameter. It and the exponential come out close, and both are far ahead of linear: the curve bends upward. The rolling rate above shows the bend getting steeper, not flattening.
Model comparison
| Model | Form | R² | ΔAIC |
|---|---|---|---|
| Exponential | A·e^kt | 0.957 | 1.6 |
| Quadratic | a + bt + ct² | 0.962 | 0.0 |
| Linear | a + bt | 0.827 | 47.8 |
ΔAIC under 2 is a tie; over 10 rules a model out. The quadratic ties the exponential because both bend upward. The straight line is ruled out (Δ ≈ 48).
Data table, all weeks
| Week of | Tokens | vs prior week | Exp. fit | Top model |
|---|---|---|---|---|
| Feb 2, 2026 | 9.80T | — | 11.0T | moonshotai/kimi-k2.5 |
| Feb 9, 2026 | 13.0T | +32.8% | 11.9T | minimax/minimax-m2.5 |
| Feb 16, 2026 | 13.9T | +7.2% | 12.8T | minimax/minimax-m2.5 |
| Feb 23, 2026 | 13.6T | −2.4% | 13.8T | minimax/minimax-m2.5 |
| Mar 2, 2026 | 14.8T | +8.6% | 14.8T | minimax/minimax-m2.5 |
| Mar 9, 2026 | 16.9T | +14.1% | 15.9T | minimax/minimax-m2.5 |
| Mar 16, 2026 | 20.3T | +20.3% | 17.2T | xiaomi/mimo-v2-pro |
| Mar 23, 2026 | 22.6T | +11.5% | 18.5T | xiaomi/mimo-v2-pro |
| Mar 30, 2026 | 27.0T | +19.3% | 19.9T | qwen/qwen3.6-plus-04-02:free |
| Apr 6, 2026 | 21.0T | −22.3% | 21.4T | qwen/qwen3.6-plus-04-02:free |
| Apr 13, 2026 | 20.5T | −2.5% | 23.0T | anthropic/claude-4.6-sonnet |
| Apr 20, 2026 | 21.9T | +6.9% | 24.8T | moonshotai/kimi-k2.6 |
| Apr 27, 2026 | 23.8T | +8.8% | 26.7T | tencent/hy3-preview:free |
| May 4, 2026 | 25.7T | +8.1% | 28.7T | tencent/hy3-preview:free |
| May 11, 2026 | 26.9T | +4.6% | 30.9T | tencent/hy3-preview |
| May 18, 2026 | 28.9T | +7.5% | 33.3T | deepseek/deepseek-v4-flash |
| May 25, 2026 | 31.8T | +10.0% | 35.8T | deepseek/deepseek-v4-flash |
| Jun 1, 2026 | 36.1T | +13.6% | 38.6T | deepseek/deepseek-v4-flash |
| Jun 8, 2026 | 44.6T | +23.5% | 41.5T | deepseek/deepseek-v4-flash |
| Jun 15, 2026 | 46.6T | +4.6% | 44.7T | deepseek/deepseek-v4-flash |
| Jun 22, 2026 | 46.7T | +0.2% | 48.1T | deepseek/deepseek-v4-flash |
| Jun 29, 2026 | 46.7T | +0.0% | 51.8T | deepseek/deepseek-v4-flash |
| Jul 6, 2026 | 52.6T | +12.6% | 55.8T | tencent/hy3:free |
| Jul 13, 2026 | 62.8T | +19.3% | 60.0T | tencent/hy3:free |
| Jul 20, 2026 | 58.0T | −7.5% | 64.6T | xiaomi/mimo-v2.5 |
| Jul 27, 2026 | 56.8T | −2.2% | 69.6T | deepseek/deepseek-v4-flash |
| Aug 3, 2026 | 69.0T | +21.6% | 74.9T | deepseek/deepseek-v4-flash-0731 |
| Aug 10, 2026 | 75.3T | +9.1% | 80.6T | deepseek/deepseek-v4-flash-0731 |
| Aug 17, 2026 | 93.4T | +23.9% | 86.8T | deepseek/deepseek-v4-flash-0731 |
| Aug 24, 2026 | 113T | +21.0% | 93.4T | stealth/ox-alpha |
| Aug 31, 2026 | 115T | +2.2% | 101T | tencent/hy4-preview |
| Sep 7, 2026 | 127T | +9.8% | 108T | openai/gpt-5.6-luna |
| Sep 14, 2026 | 129T | +1.7% | 117T | deepseek/deepseek-v4.1-flash |
| Sep 21, 2026 (partial) | 50.5T | — | 125T | z-ai/glm-5.3-flash |
Who is growing
The same platform over a full year, by model author. Only each week's top 9 models can be attributed to an author; the rest of the platform is shown as other models (35–80% of a week's tokens). Switch to shares to see the leaderboard churn.
Last week, by author
The latest complete week, in the same author colors. Anthropic, Google and MiniMax had no model in that week's top 9, so their tokens are part of other models.
Requests grow too, but slower
OpenRouter's market-share chart counts requests rather than tokens, one small chart per author on its own scale. Gaps are weeks in which an author was not in the top 9.
Bigger requests, not just more of them
Over the year, weekly tokens grew about 24× and weekly requests about 5×, so the average request grew from about 5,000 to 24,000 tokens (prompt plus completion), in line with agents that re-send long contexts at every step.
Models and tasks
Every text model OpenRouter lists, by prompt price and context window. Bubbles are the models that made a weekly top 9, sized by their tokens in their latest such week; grey dots are the rest of the catalog. Most of the volume sits on cheap models with 1M-token windows.
What the tokens are spent on, over the 30 days before Sep 23: agent workflows and coding lead.
Beyond OpenRouter
Token volumes that providers have announced, converted to tokens per day. The series measure different things (all of Google's products, API traffic only, China's national total), so compare their slopes, not their levels. OpenRouter is the thick red line.
- Google (all products) · trillion tokens per month, 5 figures: blog.google/…/sundar-pichai-io-2026, blog.google/…/alphabet-earnings-q2-2025, blog.google/…/gemini-enterprise-sundar-p…
- Google (customer API) · billion tokens per minute, 5 figures: blog.google/…/alphabet-earnings-q3-2025, blog.google/…/alphabet-earnings-q4-2025, blog.google/…/cloud-next-2026-sundar-pic…, blog.google/…/sundar-pichai-io-2026, blog.google/…/alphabet-earnings-q2-2026
- OpenAI (API) · billion tokens per minute, 2 figures: techcrunch.com, constellationr.com
- Microsoft (Azure AI Foundry) · trillion tokens per month, 3 figures: microsoft.com/…/earnings-fy-2025-q4, microsoft.com/…/earnings-fy-2025-q3
- OpenRouter (platform) · trillion tokens per week, 7 figures: cryptobriefing.com, the-decoder.com, a16z.com, siliconangle.com, techcrunch.com
- ByteDance (Doubao / Volcano Engine) · trillion tokens per day, 8 figures: caixinglobal.com, silicon.co.uk, finance.yahoo.com, en.tmtpost.com, news.aibase.com, global.chinadaily.com.cn, finance.biggo.com
- China (national total) · trillion tokens per day, 6 figures: ecns.cn, globaltimes.cn, technode.com/…/china-reports-surge-in-tok…, technode.com/…/chinas-daily-ai-token-usag…
- Fireworks AI · trillion tokens per day, 3 figures: fireworks.ai/…/series-c, io-fund.com, fireworks.ai/…/series-d-announcement
Sources
OpenRouter data: openrouter.ai/rankings (Top Models, Market Share and task charts) and the public models API, retrieved Sep 23, 2026 (usage data through Sep 22). It covers traffic routed through OpenRouter only. Industry figures: the providers' own posts and earnings materials, or press reports of them, listed above. Chart sources: examples/demos/openrouter.