AI usage · October 2026
AI use is up 330×. The bill is up 115×.
In 21 months, weekly AI use on OpenRouter went from half a trillion tokens to 164 trillion. And it’s speeding up: in 2025 it doubled every 103 days, this year every 59.
OpenRouter sells access to hundreds of AI models through one account, and it publishes how many tokens, the chunks of text a model reads and writes, go to each one every day. In the week of 6 Jan 2025 that was 0.5 trillion. In the week of 12 Oct 2026 it was 164 trillion: 330 times as much.
OpenRouter doesn’t publish what people pay, so we priced every model’s tokens at its list price. The weekly bill grew about 115×, from roughly $1.5 million to $176 million.
Tokens are input plus output, all models. Spend is each model’s tokens times its list price, $0 for free models. Full weeks from 6 Jan 2025. Hover or tap for each week.
And it’s speeding up
In 2025, weekly tokens grew 11×, doubling about every 103 days. In the 9 months of 2026 so far they have grown 29×, doubling every 59 days. Going from 1 to 10 trillion tokens a week took 50 weeks. Going from 10 to 100 trillion took 28.
The turn came early this year. OpenRouter estimates that around 6 February 2026, AI agents, programs that call a model over and over to finish a task, passed people typing into chat windows in daily token use. A person reads a few answers a day. An agent working through a codebase can send millions of tokens an hour.
Counted from the first full week at or above each level to the first at or above ten times it.
The bill is catching up
Spending grew more slowly because most of the new tokens went to cheap models. Closed models, mostly from OpenAI, Anthropic and Google, handled 73% of tokens in January 2025 and 24% in Sep 2026; the rest went to open models, mostly from Chinese labs, to free ones and to models outside the top 50. So the average token fell from $3.06 per million to $1.07 (more on AI prices).
But the gap is closing. In 2025, tokens grew 11× and spending 5.3×. This year it’s 29× and 22×. The average token still gets cheaper, but more slowly: its price fell 52% in 2025 and 27% so far this year.
Vercel sees the same thing
OpenRouter is one marketplace, popular with coding agents and with people trying free models. Vercel’s AI Gateway, which serves mostly companies’ apps, is another, and its own report shows the same climb. From December 2025 to September 2026, spending on its gateway grew about 13× while the average price per token fell 47%, so tokens grew about 25× in nine months.
Cheap tokens didn’t shrink the bill
AI use isn’t growing like an app people sign up for. It compounds, because the fastest-growing users are programs, not people, and a program uses as many tokens as it can afford. As the price fell, agents used more. This year the average token got 27% cheaper, and spending still grew 22×.
The second tenfold step took about half as long as the first. If this year’s pace held, weekly use on OpenRouter would pass 1,000 trillion tokens around Mar 2027. A pace like that won’t last forever, but so far nothing in the data says it’s slowing down.
Method
Tokens are from OpenRouter’s Data API (rankings-daily, CC BY 4.0): every day from 1 January 2025 to 23 Oct 2026, the input plus output tokens of the top 50 models, plus one row for all the others (Source: OpenRouter (openrouter.ai/rankings), as of 2026-10-24.). Weeks run Monday to Sunday and only full weeks count; OpenRouter’s data skips two days (15 June and 15 July 2025), so those weeks are scaled up from six days. Doubling times are averages between the first full week of 2025, the last week starting in 2025 and the latest full week.
Spend is our estimate, not what anyone paid. Each model’s daily tokens are priced at its list price on OpenRouter as of 24 Oct 2026, weighted 3 input tokens to 1 output token (the mix Artificial Analysis uses), with the earlier prices of models whose price later changed (for example GPT-5.6 Luna and DeepSeek V4 Pro) and public list prices for retired models. Free models and unreleased test models count as $0. Models outside the top 50 get that day’s average price. Real bills are lower: list prices ignore cached input, which coding agents use heavily and costs a tenth as much, and temporary discounts. Weighting 9 input tokens to 1 changes the growth in spending from 115× to about 116×. Closed means OpenRouter lists no published weights for the model.
Vercel’s figures are from its AI Gateway production index for September 2026; Vercel publishes changes and shares, not totals. The data and code are in our repository (data/token-economics).