OpenRouter was handling over 1 trillion tokens a day by late 2025
Published 2026-11-12
Do cheaper tokens mean a smaller AI bill? List prices per million tokens sit next to OpenRouter's usage growth, with one test you can run this week.
By late 2025, OpenRouter was handling over 1 trillion tokens a day. If your team pays per token, volume sets your bill, not the price list.
a16z's State of AI reports OpenRouter's token volume. OpenRouter routes developers' requests to AI models from many providers, so its volume is one platform's read on AI use. A token is a small chunk of text, a few letters or part of a word, that an AI model reads or writes. Vendors bill by the token.
Your bill is price times volume, so watch both.
What changed
On volume: a16z says OpenRouter grew from handling roughly 10 trillion tokens per year to more than 100 trillion as of mid-2025. In late November 2025 it processed more than 1 trillion tokens every day. Source: a16z, State of AI.
On price: OpenAI's launch post for its GPT model family prices it per million tokens. Sol is $5 input and $30 output, Terra is $2.50 and $15, and Luna is $1 and $6. Source: OpenAI.
Google prices Gemini Decision 1 at $0.10 per million input tokens and $0 for output tokens. Source: Google.
The money
You pay the model maker twice per request: once for the tokens you send in, once for the tokens that come back. Both are priced per million.
| Model | Input per million | Output per million |
|---|---|---|
| OpenAI Sol | $5 | $30 |
| OpenAI Terra | $2.50 | $15 |
| OpenAI Luna | $1 | $6 |
| Google Gemini Decision 1 | $0.10 | $0 |
Inside one OpenAI family, Sol's price is five times Luna's on both input and output.
Going by the listed prices, Gemini Decision 1's input price is one tenth of Luna's. The two may not do the same job, so treat it as a price gap, not a like-for-like saving.
Google lists $0 for output on Gemini Decision 1, so input is the only priced line there.
So what?
Your spend is price times volume. List prices are a menu you can pick from. Volume is what you control, and it is what grew on OpenRouter: past 1 trillion tokens a day by late 2025.
Every workflow you add to your stack adds to your own volume. Each one looks small alone. Together they set the bill.
So the question for your next budget review is not only what a token costs. It is also how many tokens your team may send if AI becomes cheap enough to use widely.
The move
Who it's for: an operator or product lead whose team runs a high-volume AI step, such as tagging, routing, or first-pass review.
The job: keep cost per task flat while usage climbs.
Three moves to test:
- Route by difficulty. Send simple, repetitive work to the cheapest tier and keep the top tier for hard cases. Luna's input price is a fifth of Sol's.
- Renegotiate by tier. At renewal, ask your vendor for per-million prices on every tier, input and output separately, and for what changes when your volume doubles.
- Founders: price your product per task, not per token. Your cost scales with volume, so a flat seat price can get eaten by your heaviest users.
The catch
The Jevons paradox is the general idea that when a resource gets cheaper to use, people can use so much more of it that total spend goes up. The sources do not show that it explains OpenRouter's growth.
They show usage rising. They do not show that falling prices caused it, or that anyone's total bill went up. OpenRouter's growth could come from new users and new apps as much as from cheaper tokens.
One platform is also not the market.
And a low price only helps if the cheap tier does the job. If it fails more often, redoing the work can cost more than the top tier would have. That is why the test below measures quality as well as cost.
Try this
Find your team's highest-volume AI task.
- Pull its cost and call count for the last month from your vendor invoice or usage dashboard. Divide to get cost per task.
- Take a sample of real inputs and rerun them on the cheapest tier your vendor offers.
- Have someone on your team grade both sets of outputs against the same checklist.
- Compare cost per task and pass rate.
It worked if the cheap tier passes at a rate you would accept and cost per task drops. If quality falls short, you have your reason to keep the top tier, with numbers behind it.
Prices come from the launch posts linked above and may have changed since. The ideas here are untested.
All source links
- a16z.com/state-of-ai checked 12 November 2026.
- blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api checked 12 November 2026.
- openai.com/index/gpt-5-6 checked 12 November 2026.
