Skip to content
Finitizer

AI token usage by model, 30 days to 13 Nov 2026

OpenRouter processed 834 trillion AI tokens in the 30 days to 13 Nov 2026, across 589 models.

DeepSeek V4.1 Flash 1017 alone took 22.2% of them, and the ten most used models took 61.3%. OpenRouter is a marketplace that routes requests to models from many labs, so its traffic shows which models developers pick. It is one marketplace, not the whole industry.

Usage read 14 Nov 2026Next update 1 Dec 2026Source: OpenRouter, CC BY 4.0
Tokens in 30 days
834T
to 13 Nov 2026
Tokens in the last 7 days
210T
to 13 Nov 2026
Share of the top model
22.2%
DeepSeek V4.1 Flash 1017
Share of the top 10 models
61.3%
10 of 589 models

T is trillions of tokens, B is billions. Tokens are prompt plus completion.

Which models the tokens went to

Share of all tokens in the 30 days to 13 Nov 2026. The ten most used models, then every other model as one row, so the bars add up to the whole.

  • DeepSeek V4.1 Flash 1017DeepSeek22.2%185T
  • DeepSeek V4.1 FlashDeepSeek9.5%79.4T
  • GLM 5.3 FlashZ.ai6.9%57.4T
  • GPT-6 LunaOpenAI4.2%34.8T
  • Claude Opus 5.5Anthropic3.7%30.9T
  • MiniMax M3.1 FlashMiniMax3.4%28.0T
  • Hy4Tencent3.2%26.9T
  • Hy4 previewTencent3.1%26.0T
  • Step 5 PreviewStepFun2.8%23.2T
  • MiMo-V2.6-FlashXiaomi2.4%19.7T
  • All other models (579)38.7%322T

The 25 most used models

The 25 AI models that processed the most tokens on OpenRouter in the 30 days to 13 Nov 2026: tokens, share of all tokens, tokens in the last 7 days and tokens per request.
#Model30 daysShare
1DeepSeek V4.1 Flash 1017DeepSeek185T22.2%
2DeepSeek V4.1 FlashDeepSeek79.4T9.5%
3GLM 5.3 FlashZ.ai57.4T6.9%
4GPT-6 LunaOpenAI34.8T4.2%
5Claude Opus 5.5Anthropic30.9T3.7%
6MiniMax M3.1 FlashMiniMax28.0T3.4%
7Hy4Tencent26.9T3.2%
8Hy4 previewTencent26.0T3.1%
9Step 5 PreviewStepFun23.2T2.8%
10MiMo-V2.6-FlashXiaomi19.7T2.4%
11GLM 5.3Z.ai19.0T2.3%
12DeepSeek V4 Flash 0731DeepSeek16.2T1.9%
13Nemotron 3 UltraNVIDIA16.0T1.9%
14GPT-6.1 SolOpenAI14.6T1.7%
15Jev 1.13TypeSafe12.8T1.5%
16Muse Spark 1.3 ContributorMeta12.2T1.5%
17Hy3Tencent11.9T1.4%
18Claude Haiku 5.5Anthropic11.6T1.4%
19Nemotron 3.5 Super VLNVIDIA11.3T1.4%
20GPT-5.6 LunaOpenAI11.1T1.3%
21DeepSeek V4 Flash 0423DeepSeek11.0T1.3%
22Claude Sonnet 5.5Anthropic10.7T1.3%
23Kimi K3.1 ProMoonshot AI10.0T1.2%
24Kimi K3Moonshot AI7.4T0.9%
25Jev 1.14TypeSafe7.4T0.9%

A dated name such as 1017 is a snapshot of a model, which OpenRouter ranks apart from the model's other snapshots. Tokens per request is high where agents resend a long context on every turn.

Which labs made those models

DeepSeek models took 36.0% of all tokens, ahead of Z.ai at 9.8% and OpenAI at 9.5%.

  • DeepSeek15 models36.0%300T
  • Z.ai17 models9.8%81.8T
  • OpenAI80 models9.5%79.5T
  • Tencent10 models7.8%64.8T
  • Anthropic17 models7.6%63.1T
  • NVIDIA15 models3.9%32.8T
  • MiniMax13 models3.8%32.0T
  • Google42 models3.4%28.2T
  • Xiaomi5 models3.1%26.2T
  • StepFun3 models2.9%24.1T
  • All other labs12.2%101T

Where the labs are based

Models from labs based in China took 67.5% of the tokens, and models from labs based in the United States 27.5%.

  • Labs based in China67.5%564T
  • Labs based in the United States27.5%230T
  • Labs based elsewhere1.3%10.6T
  • Not classified3.7%30.5T

The country is where the lab that makes the model is based, classified by Finitizer. It says nothing about where a request came from or which servers ran it. China here is DeepSeek, Z.ai, Tencent, MiniMax, Xiaomi and StepFun and 6 smaller labs. Not classified covers stealth models, labs with no public home, and the labs too small to be named in the reading.

Tokens are not spend

A ranking by tokens favors cheap models. Read these four figures beside the charts above before using token share as a guide to where the money goes.

  • Over 70% of tokens, about 25% of spend

    What open-weight models took of paid OpenRouter traffic: most of the tokens, about a quarter of the money. October 2026. OpenRouter

  • $1.21 to $0.52, and $0.28 to $0.084 per million tokens

    Blended price per million tokens on OpenRouter for closed-weight models, then for open-weight models: down 57% to 70%. January to October 2026. OpenRouter

  • 76% of tokens from agents

    Share of OpenRouter tokens used by API keys it classes as agents, up from 8% a year earlier. October 2026. An agent resends its context on every step, so it uses many times the tokens of a chat. OpenRouter

  • 6.6% of tokens on free variants

    Share of the 30 days to 13 Nov 2026 that ran on the free variant of a model, where the user paid nothing per token. From this reading.

For a budget, the mix matters more than the total: the same workload costs several times more or less depending on the model, how much of the prompt is cached, and how many steps an agent takes. Price a workload on the LLM Cost Calculator, or compare list prices in the LLM price index.

Is your own token bill growing faster than your usage?

Finitizer Tokenomics: Reads your AI bills and exports and shows what changed, who spent it, and what to fix. No proxy, no code change. Founding customers list now open.

See Tokenomics

How fast usage has grown

Tokens OpenRouter processed in a week: three figures it has published, then the 210 trillion recorded here for the 7 days to 13 Nov 2026.

Tokens processed on OpenRouter in a week, Jan 2025 to Nov 2026: Jan 2025 0.5T, May 2026 25T, Sep 2026 128.4T, Nov 2026 210.2T.Jan 2025: 0.5T tokens a week0.5TJan 2025May 2026: 25T tokens a week25TMay 2026Sep 2026: 128.4T tokens a week128.4TSep 2026Nov 2026: 210.2T tokens a week210.2TNov 202620252026

Rings are figures OpenRouter published; dots are readings recorded on this page. The line between them is a guide, not a measurement.

Readings recorded here

This is the first reading. The change over 1 month, 3 months, 6 months and 12 months appears here as later months are recorded; nothing is estimated in the meantime.

Every monthly reading of AI token usage on OpenRouter recorded on this page, newest first.
30 days toTokens, 30 daysTokens, last 7 days
13 Nov 2026834T210T

What OpenRouter and Google have published

Published figures on AI token usage, each with the period it is for and its source.
PeriodFigureSource
Week of 13 Jan 20250.5 trillion tokensOpenRouter, The Token State of Play
Late May 202625 trillion tokens a weekOpenRouter funding announcement
Week of 14 Sep 2026128.4 trillion tokensOpenRouter, The Token State of Play
Mid-September 202669% of tokensOpenRouter, The Token State of Play
October 2026790 trillion tokensOpenRouter, October 2026 data recap
October 202676% of tokensOpenRouter, October 2026 data recap
October 2026Over 70% of tokens, about 25% of spendOpenRouter, October 2026 data recap
January to October 2026$1.21 to $0.52, and $0.28 to $0.084OpenRouter, October 2026 data recap
Third quarter of 2026Nearly 32 billion tokens a minuteAlphabet third-quarter 2026 earnings call, 28 Oct 2026

OpenRouter is a small part of all AI use. At nearly 32 billion tokens a minute, Google's own API alone processes about 1.38 quadrillion tokens in 30 days, by our arithmetic from its stated rate.

Keep these figures

Have this month's figures sent to you, and choose whether to hear when the price pages on this site change.

How this is measured

  • Read once a month from OpenRouter's rankings data: tokens by model over the 30 days and the 7 days before the read.
  • Tokens are prompt plus completion tokens, counted by each model provider's own tokenizer, so a token of one model is not exactly a token of another.
  • The standard, free and batch variants of a model are added together. Each dated snapshot of a model is counted apart, as OpenRouter ranks them.
  • Only public traffic through OpenRouter is counted. Use of a model directly from its maker, or through a cloud, is not here.
  • A lab is the maker of the model. Its country is where that maker is based, classified by Finitizer from public information.
  • A change compares this reading with the one taken about that many months earlier. Where there is none, the page says no data.
  • A recorded month is never rewritten.

Sources, read 14 Nov 2026

Use the data

Every reading is free to download as CSV or JSON, under CC BY 4.0, the license OpenRouter gives its rankings data. Credit it as: Source: OpenRouter (openrouter.ai/rankings), as of 14 Nov 2026. Licensed under CC BY 4.0. Tables by Finitizer (tools.finitizer.com/ai-token-usage).

Source: OpenRouter (openrouter.ai/rankings), as of 14 Nov 2026. Licensed under CC BY 4.0.

Questions people ask

How many AI tokens does OpenRouter process?

834 trillion tokens in the 30 days to 13 Nov 2026, and 210 trillion in the last 7 of those days. OpenRouter itself reported 790 trillion tokens for October 2026. A token is about three quarters of an English word, and the count is prompt plus completion tokens.

Which AI model is used the most?

By tokens on OpenRouter in the 30 days to 13 Nov 2026, DeepSeek V4.1 Flash 1017 from DeepSeek: 185 trillion tokens, 22.2% of the total. The ten most used models took 61.3%. This ranks use on one marketplace, not quality.

How much AI token usage goes to Chinese models?

On OpenRouter in the 30 days to 13 Nov 2026, models from labs based in China took 67.5% of tokens and models from labs based in the United States took 27.5%. The country is where the lab that makes the model is based, classified by Finitizer; it says nothing about where a request came from or where it ran.

Is token usage the same as AI spending?

No. A cheap model can lead on tokens and trail on money. OpenRouter reported that open-weight models took over 70% of tokens, about 25% of spend in October 2026, and that its blended price per million tokens fell from $1.21 to $0.52 for closed-weight models and from $0.28 to $0.084 for open-weight models over January to October 2026. Tokens grow much faster than spend.

Does this show the whole AI market?

No. It is one marketplace's public traffic. Most use of a model goes straight to its maker or through a cloud, and is not counted here: Google said its own API handles nearly 32 billion tokens a minute, which is about 1.38 quadrillion tokens in 30 days. OpenRouter's mix also leans toward developers who compare models and prices.

How often is this page updated?

Once a month. This reading was taken on 14 Nov 2026 and covers usage to 13 Nov 2026; the next is due on 1 Dec 2026. A recorded month is never rewritten, so the history can be compared month to month.

Market share is their tokens. What about yours?

Finitizer Tokenomics shows what your AI actually costs, read from your own bills and exports: by workload, team, model and the key carrying it.

See Finitizer Tokenomics

Related