AI Token Usage in 2026: Google Hits 3.2 Quadrillion a Month, OpenRouter Grows 250x, and What It Means for Your AI Budget

Tokens have become the electricity meter of the AI economy. Every question to a chatbot, every document an agent reads and every line of code it writes is billed in tokens, and the volumes are growing faster than almost any other technology metric. Google now processes more than 3.2 quadrillion tokens a month, seven times more than a year ago. OpenAI's API more than doubled its throughput in five months, and traffic on the OpenRouter marketplace grew 250-fold in 20 months. This article collects the published numbers, explains why agents are behind the jump, and shows what it means for the AI budget of a company in Croatia or the DACH region.

Why tokens are the metric to watch

A token is the unit a language model reads and writes: in English, roughly three quarters of a word. Croatian and German usually need more tokens for the same text, because tokenizers are trained mostly on English. Providers bill input and output tokens separately, so token volume is the closest thing the industry has to a usage meter. User numbers say how many people open an app; tokens say how much work the models actually do.

The figures below come from the companies themselves (keynotes, earnings releases) and from OpenRouter, a marketplace that routes developer requests to hundreds of models and publishes its traffic. None of them is audited, and each company counts differently. They still show the order of magnitude and the direction very clearly.

Google: from 9.7 trillion to 3.2 quadrillion tokens a month

Google is the only company that has published a consistent series. At I/O 2024 it processed 9.7 trillion tokens a month, at I/O 2025 about 480 trillion, and at I/O 2026 more than 3.2 quadrillion: 7 times more in one year and about 330 times more in two. The number covers all surfaces, including Search with AI Overviews and AI Mode, the Gemini app, Workspace and the developer APIs.

Other figures from the same keynote show where the growth comes from. The Gemini app grew from 400 million to more than 900 million monthly users (950 million in the Q2 results), but its daily requests grew more than seven times, so usage per user is rising faster than the user base. More than 375 Google Cloud customers each processed over one trillion tokens in the past twelve months, and over 8.5 million developers build with Google's models every month. Inside Google, AI developer tools went from about half a trillion to more than three trillion tokens a day between March and May.

API throughput: OpenAI and Google side by side

The API numbers measure what developers and companies build, not consumer chat. OpenAI reported 6 billion tokens per minute at DevDay in October 2025 and, according to reports from March 2026, more than 15 billion, about 2.5 times more in five months. Google reported more than 16 billion per minute for its first-party models in the Q1 2026 results, about 19 billion at I/O and 22 billion in the Q2 results in July. Converted to a month, 22 billion per minute is roughly 950 trillion tokens, close to one quadrillion.

Anthropic does not publish token volumes, and neither do Mistral and Manus. Microsoft last gave a number for its Foundry APIs in July 2025: over 500 trillion tokens in fiscal year 2025, up more than 7 times. For Anthropic the best available proxies are revenue and enterprise spend data, which we cover in enterprise LLM market share 2026.

OpenRouter: 250x in 20 months, and agents now use more tokens than people

OpenRouter is a useful window because it shows all labs on one platform. Its weekly volume went from 0.5 trillion tokens in January 2025 to 128.4 trillion in mid-September 2026, about 250 times more, and an analysis of its data puts the week ending 5 October at almost 168 trillion. OpenRouter estimates that 6 February 2026 was the last day on which people used more tokens than agents; since then automated workflows have consumed the majority.

The mix has changed just as much. Open-weight models first passed 50% of tokens in February 2026, peaked at 79% in early July and have stayed above 62% since. Models from Chinese labs (DeepSeek, Moonshot, Z.ai, Qwen and others) handled 69% of tokens in the latest week covered, while labs outside the US and China, the group that includes Mistral, fell from almost 12% of tokens at the start of 2025 to about 2%. Between the two American frontier labs, OpenAI took 78% of combined OpenAI and Anthropic tokens in mid-September and, for the first time in this series, just over half of their combined spend.

This is the most useful lesson for buyers. Open-weight models process most of the tokens because they are cheap enough for high-volume steps such as classification, extraction and sub-agent work. Closed frontier models from OpenAI, Anthropic and Google still take almost three quarters of the money, because companies pay for them where quality decides the result. Keep in mind that OpenRouter is a small, developer- and start-up-heavy slice of the market; enterprise traffic that goes directly to Anthropic, OpenAI, Azure, Bedrock or Vertex AI is not in these numbers.

Why the volume explodes

Three effects multiply each other:

• Agents. A chatbot answer is one model call. An agent that fixes a bug, researches a market or processes an invoice makes dozens of calls, re-reads its growing context at every step and calls tools in between. One task can easily consume a hundred times more tokens than one chat answer. See our guide to context engineering for agents.
• Reasoning models. Models that "think" before answering generate thinking tokens that are billed as output, often several times longer than the answer itself.
• Falling prices. Epoch AI found that the price of reaching a given benchmark score fell by a median of about 50x per year (between 9x and 900x depending on the task); a NeurIPS 2025 workshop paper estimates 5x to 10x per year for frontier models. Every price drop makes new use cases economical, so total spend rises even as the price per token falls. Economists call this the Jevons paradox.

What it means for companies in Croatia and the DACH region

The growth curves also show that AI usage is concentrating where it is cheapest to run at scale. For a mid-sized company in Zagreb, Vienna or Munich this is good news: the capability that cost a fortune two years ago is now cheap enough for everyday processes. The risk is the opposite of the usual one: not that AI is too expensive to try, but that agents quietly consume a budget nobody planned for. Companies that measure cost per task from the first pilot keep that under control.

How reliable are these numbers?

Treat all of them as indicators. The company figures are self-reported and not audited, and Google's numbers mix consumer, search and API traffic. OpenAI's 15 billion per minute comes from reports citing the company, not from an OpenAI release we could find. OpenRouter covers only traffic that runs through its platform, which is skewed towards developers, start-ups and price-sensitive users and excludes resellers. Different trackers of the same platform also disagree: one brokerage report put Chinese models at 65% of OpenRouter tokens in September, OpenRouter itself at 69%. The direction, however, is the same in every source: AI usage measured in tokens is growing by multiples per year, not by percent.

FAQ

How many AI tokens does Google process per month?

More than 3.2 quadrillion across all AI products (Google I/O, May 2026), about seven times more than the roughly 480 trillion of May 2025. Gemini models process 22 billion API tokens per minute (Q2 2026).

How many tokens does the OpenAI API process?

6 billion tokens per minute at DevDay in October 2025, and more than 15 billion per minute according to reports from March 2026.

Does Anthropic publish how many tokens Claude processes?

No. On OpenRouter, OpenAI took 78% of combined OpenAI and Anthropic tokens and just over half of the spend in September 2026, but OpenRouter is a small, developer-heavy slice of the market.

Why is AI token usage growing so fast?

Agents make many model calls per task, reasoning models produce thinking tokens, and the price for the same capability falls by a median of about 50x per year (Epoch AI), which makes many more tasks worth automating.

How should a company budget for AI tokens?

Per completed task, not per token: measure in a pilot, cache repeated context, route simple steps to cheaper models, batch what can wait and cap agent loops.