OpenRouter at 10 Quadrillion Tokens a Year? What the Claim Says and What the Public Data Shows

A post on X says OpenRouter has crossed a 10 quadrillion annual token run rate. The investor Deedy Das wrote on November 2, 2026: "Congratulations to OpenRouter for crossing 10 QUADRILLION annual token run rate today! Token volumes continue to 2x every 10.5 weeks. In the last 9 weeks, OpenRouter processed more tokens than all of time before that." OpenRouter's CEO, Alex Atallah, reposted it with a line of his own: "The engineering that goes into scaling OpenRouter is beyond anything I've seen before in my career."

We could not find an OpenRouter page that states the total. This post checks the claim against what OpenRouter does publish, its public rankings, and against basic arithmetic. The short answer: the order of magnitude holds up, the exact figure is unconfirmed, and the two growth claims do not line up perfectly.

Claim, evidence and status

What does 10 quadrillion tokens a year mean?

Ten quadrillion is 10 to the 16th power, or 10,000 trillion tokens. Spread over a year, that is roughly 27 trillion tokens a day, or about 190 trillion a week. A run rate is not a total. It takes a recent period, here presumably a recent week or day, and multiplies it up to a year, so a company growing quickly looks bigger on a run-rate basis than on a trailing-year basis.

What do OpenRouter's public rankings show?

The rankings page says it was updated with usage data through November 1, 2026, and ranks models by tokens processed, counting prompts and completions. It shows several windows. Using the figures on the page:

• Latest complete day, top ten models: 6 trillion, 2.14 trillion, 1.41 trillion, 900 billion, 803 billion, 748 billion, 735 billion, 654 billion, 572 billion and 571 billion tokens. Adding them gives about 14.5 trillion tokens in one day from the ten largest models.
• Trailing 30 days, top ten: 114 trillion, 85.4 trillion, 55.1 trillion, 35.6 trillion, 33.1 trillion, 32.7 trillion, 30.2 trillion, 26.9 trillion, 22.4 trillion and 19.1 trillion. That sums to about 454 trillion, or about 15 trillion a day.

These sums are our own arithmetic from the page's tables. They leave out every model beyond the top ten, which the page groups as other. A 27 trillion token day would mean the top ten are around half of volume. That is plausible for a marketplace with a long tail of models, but the page does not give the total, so we cannot confirm it. Two caveats: the rankings exclude requests that users or apps keep private, and the page itself says it does not measure users or spend.

So the check is: the public data is consistent with a total in the tens of trillions of tokens per day, which fits a run rate on the order of 10 quadrillion a year. It does not prove the exact number.

Do the two growth claims agree?

Not quite, and this is the most useful part of the check. Suppose volume grows smoothly and doubles every 10.5 weeks. Then the growth rate is about 6.6 percent a week, and the tokens processed in the last 9 weeks are about 45 percent of everything processed up to now, while about 55 percent came before. In that world, the last 9 weeks would be slightly less than all earlier time, not more.

For the last 9 weeks to exceed all earlier time, the doubling time has to be under 9 weeks. Three explanations fit:

1. Growth sped up recently, so "2x every 10.5 weeks" is an average over a longer period.
2. The wording is loose, and "more than all of time before" is rounded up.
3. The run rate was measured on a short, high-volume window, not a smooth curve.

None of these is a criticism. It is a reminder that viral numbers from one post mix a few separate measures, and each needs its own source.

How should you read a run-rate number?

Run-rate figures are common in AI infrastructure news, and they are easy to misread. Four questions help.

What window was multiplied? A run rate built from one peak day is higher than one built from a weekly average. The post does not say which one it uses. OpenRouter's own rankings show that day-to-day swings are large: the top model moved by 15 percent in a day and some models moved by more than 100 percent.

Whose tokens are counted? OpenRouter counts prompt and completion tokens together, for requests that users and apps have not marked private. A long agent session with large cached prompts adds many prompt tokens. A figure that counts both is bigger than one that counts only generated output, and neither is a measure of value or revenue.

Is it requests, tokens or spend? The page is explicit that it measures none of users or spend. A model that is cheap and chatty can top a token chart while earning little. That is why the market-share table uses requests while the model table uses tokens.

Who benefits from the number? An investor congratulating a portfolio company and a CEO resharing it both have a reason to like a large figure. That does not make it wrong. It means the standard to apply is a published definition and a dataset, which this claim does not yet have.

Who is using it? What the rankings do show

The rankings page also gives a picture of who is driving volume, and this part is public and readable:

• Market share by author, week beginning Oct 26, 2026 (share of text requests): DeepSeek 29.8 percent, OpenAI 18.8 percent, Google 14.4 percent, Z.ai 11.9 percent, Anthropic 5.7 percent, Qwen 5.6 percent, Mistral AI 2.2 percent, Tencent 1.9 percent and Meta Llama 1.5 percent. This counts requests, not tokens or spend.
• Top model by tokens this week: DeepSeek V4.1 Flash 1017, with 48.1 trillion tokens, up 43 percent. It is an open-weights DeepSeek release we covered.
• Other fast risers: GLM 5.3 Flash, GPT-6 Luna, Step 5 Preview, MiniMax M3.1 Flash (up 690 percent week over week) and Claude Opus 5.5 (up 841 percent over 30 days from a small base).
• Top apps by tokens today: Hermes Agent at 1.67 trillion, Claude Code at 982 billion, Kilo Code at 651 billion, OpenClaw at 573 billion, pi at 470 billion, Codex at 408 billion and DeepSeek Harness at 302 billion.

The pattern is that cheap, fast models and agent harnesses carry much of the token volume. That fits what we have seen elsewhere, including our look at cheap Chinese open models and the infrastructure reset and the study of how Chinese models perform across languages.

A reminder from OpenRouter's own page: "a higher token total shows how much a model is used, not which model is best for a task." Agents generate long chains of tokens, so a few heavy agent users can move the totals.

What the model-diversity claim adds

OpenRouter's own account says companies and users spending 100 dollars or more a month now use a median of six model families, double a year ago, and the CEO argues that as models improved, diversity in the customer base went up. That is a claim about its customers, from the vendor, and we could not check the data behind it. It fits a trend we have tracked: teams route different tasks to different models. See our coverage of OpenRouter's new default routing that favours quality over price and the price-cut timeline.

What this means for what you build or pay

• Do not read token share as quality. Pick models by your own evaluations, as we argue in our Particle-1 claimed-vs-verified post, where a token-efficiency claim needed independent checking.
• Plan for multi-model routing. If your peers run six model families, a single-vendor lock is a cost risk. Keep prompts portable and test fallbacks.
• Watch for capacity and price pressure. Fast demand growth strains providers. Plan for rate limits and use caching where it helps.
• Treat marketplace stats as marketing until published. A run-rate number from a post is not a filing. Ask for the dataset or the page.

How you can check it yourself

1. Open the OpenRouter rankings and note the date of the newest usage bucket.
2. Sum the daily tokens for the top models, then compare with 27 trillion a day.
3. Look for a published total or a Data API field, since OpenRouter says rankings data is licensed CC BY 4.0 and available through a Data API.
4. Ask OpenRouter for the underlying total and the definition of run rate. We will update this post if the company publishes it.

Related reading

• DeepSeek V4.1 Flash 1017: open weights, benchmarks and pricing

• OpenRouter default routing: quality over price

• Cheap Chinese open models and the infrastructure reset

• How tokens got cheaper: the price-cut timeline

• Particle-1 on OpenRouter: claimed vs verified

• Vals AI: Chinese models and language

• Sources: OpenRouter rankings, Alex Atallah on X

• Menlo Ventures: agents generate 80% of OpenRouter tokens, Anthropic at 55% of enterprise API spend

Figures are accurate as of November 2, 2026. Rankings change daily; our sums are our own arithmetic from the page.