AI agents use roughly five times more tokens than humans
Data compiled from the OpenRouter routing layer shows AI agents consuming roughly five times more tokens per task than humans typing into the same models. Agent usage crossed above human usage on February 6, 2026 and has grown about 14 fold since.
The load has moved from chat to agents
Per PPC Land reporting dated August 25, 2026, measured data from the OpenRouter routing layer shows AI agents consuming roughly five times more tokens per task than human users typing into the same models.
In numbers: agents consume 7.3 trillion tokens on a seven day average. Per OfficeChai reporting dated August 23, human consumption at the same scale sits around 1.5 trillion.
The data was compiled by Peter Walker and circulated by Moses Sternstein through the a16z Charts of the Week newsletter on August 21, 2026. Because OpenRouter sits as a routing layer in front of many commercial models, it gives a cross section of mixed usage rather than one provider.
When the line was crossed and where the burn comes from
The date given for the crossover is precise: agent usage first rose above human usage on February 6, 2026.
Since then the agent side has grown about 14 fold in roughly six months. Over the same window human consumption rose from roughly 0.5 trillion to 1.4 trillion, a 2.8 fold increase. That is why the gap between the two curves keeps widening.
The most striking finding is the composition of that burn. More than 85 percent of agentic token consumption originates in cached prompts. Agents spend most of their compute re-reading context they have already been given rather than generating new output.
The reason is how agents work. Per the reporting, agents reload tool definitions, system prompts, memory and schema constraints across turns. The more steps a task takes, the more times the same context is read.
The distribution across companies is uneven too. Per the data, top decile enterprises output 8 times more tokens than a typical enterprise. The standout sector example is legal, where Codex adoption has risen 108 fold since February.
As context, OfficeChai also cites Cloudflare data showing bots account for more than 57 percent of all HTTP requests. Read together, the two measurements describe an internet whose center of gravity is shifting from people to machines.
AI cost is now set by architecture, not by question count
The assessment in this section is ours. The data can read as a cost warning, but what it really describes is architecture.
A business planning an AI budget usually asks how many questions the team will ask per month. This data shows that is the wrong question. Once an agent is in the loop, consumption detaches from the number of prompts a person writes and attaches to the number of steps in the task.
The second point is what the caching finding means in practice. If 85 percent of the burn comes from re-reading context, the way to cut cost is not doing less work, it is shrinking the context the agent carries. Unnecessary tool definitions, bloated system prompts and long memory records appended every turn go straight onto the invoice.
The third is measurability. Cost per task needs to be known before an agent goes to production. Otherwise a prototype that works can become unsustainable once volume arrives.
The fourth is the side effect of this trend. As agent traffic grows, data, privacy and monitoring questions get heavier on the provider side, and the OpenAI safety layer that preserves zero data retention was a product of exactly that pressure. Cost and privacy are two faces of the same curve.
What it means for businesses in Türkiye
The assessment below does not appear in the sources, it is our reading. The data contains no Türkiye specific breakdown.
Because AI cost in Türkiye is tied to the exchange rate, an uncontrolled rise in token consumption is felt more sharply here. A monthly invoice that looks acceptable abroad can be a very different line item for the same business in Türkiye.
The practical conclusion is this: the first thing to set up in an agent project is not model choice, it is a cost ceiling. Going to production without a maximum token count per task, a daily call limit and a stop threshold that triggers on an unexpected spike is risky.
The second point is scope discipline. The most common mistake we see in the field is giving an agent every document and every tool just in case. That creates an enormous context re-read every turn. Focusing the agent on a narrow task lowers both cost and error rate.
The third is picking the right work. Agents pay for themselves on repetitive, rule bound and high volume work: preparing quotes, tracking orders, handling the first tier of customer questions. Handing a task performed a few times a month to an agent is usually not economical.
The fourth is a measurement habit. After an agent is live, cost needs to be watched weekly rather than monthly, because per this data the growth is not linear.
The UNALSOFT take
In Agentic AI projects we treat token cost as a design constraint from the start, not as a surprise that shows up later. In practice that means keeping the context given to the agent as small as possible, simplifying tool definitions, calling memory when needed instead of carrying it every turn, and writing down a token budget per task. This data confirms that discipline for us. If 85 percent of consumption comes from re-reading context, the saving is not in a cheaper model, it is in freeing the agent from reading things it does not need. A cheap model does not rescue an expensive architecture, while a lean architecture can carry even an expensive model.
Do you know your agent cost per task?
Let us set the token budget and the stop threshold together.