OpenRouter: AI Agent Token Use Up 106x in 2026, Human Use 7x

Weekly agent token usage on OpenRouter is up 106x from the first week of January, and human token usage is up 7x over the same period, according to a chart that Peter Walker, OpenRouter’s head of insights, posted on X on 16 November. The chart’s footer reads “Source – OpenRouter token data”.

Our number: the agent multiple is about 15 times the human one (106 divided by 7 is 15.1, by our arithmetic). OpenRouter sorts its traffic by API key, and its October recap says it classifies “every API key as agentic, human or mixed with a weighted classifier of seven signals”.

Business Pill · AI Agents

An evergreen Business Engineer explainer on AI agents. It teaches the general idea only and does not discuss OpenRouter or this chart.

The key insight: OpenRouter’s token data, as charted by its head of insights Peter Walker, puts weekly agent token usage up 106x since the first week of January against 7x for human usage; the agent multiple is about 15 times the human one, by our arithmetic.

How OpenRouter Tells Agents From Humans

The seven signals, as the October recap lists them, are tool-call rate, input-to-output ratio, turns per request, cache hit rate, tools offered per request, time between requests and cancel rate. Keys are scored every month: “A score of 0.50 or more is agentic, 0.25 to 0.49 is mixed and below 0.25 is human”.

A token here is “prompt plus completion tokens”, the recap says. So the agent line counts every token sent and returned by keys that behave like software, and the human line counts keys that behave like a person typing.

OpenRouter’s Inferred newsletter of 8 October explains why agents run up the count. “An agent re-sends its whole context on every step”, it says, so in September cached prompt tokens were “just under 90% of all agentic tokens”, and output plus reasoning was around 1%. Output and reasoning were around 12% of human tokens.

Growth in weekly OpenRouter tokens by type, week of 16 November 2026 against the first week of January 2026: a
Growth in weekly OpenRouter tokens by type, week of 16 November 2026 against the first week of January 2026: agent tokens 106x, human tokens 7x, per OpenRouter token data charted by Peter Walker, OpenRouter’s head of insights, on X on 16 November 2026. The ratio of about 15 is our arithmetic; each multiple is against that type’s own January level.

What Was Posted on X

Posted by Peter Walker, head of insights at OpenRouter per his X bio, at 21:03 UTC on 16 November, this post says agent tokens are up 106x against the first week of January and human tokens up 7x. Its chart is titled “Percent growth in weekly tokens used by type vs first week of Jan 2026” and credits OpenRouter token data. The piece takes the agent and human definitions from OpenRouter’s own recap and newsletter.

What the Chart Shows

Walker’s chart is titled “Percent growth in weekly tokens used by type vs first week of Jan 2026”. It has two weekly lines, running to the week labelled Nov 16. The agent line ends at 106x and the human line at 7x. Walker’s post text reads: “Agent tokens = +106x vs first week of Jan”.

Each multiple measures one type against its own level in the first week of January. Neither says how many tokens agents and humans used. For that, OpenRouter’s newsletter says agents first used more tokens than humans on 6 February, and that “In the last week of September, agentic API keys used about 7x the tokens from human API keys”.

The monthly view points the same way. OpenRouter said agentic keys took 76% of its tokens in October, about 600 trillion of the month’s 790 trillion, against 8% a year earlier (our piece on the October recap).

How It Compares With Spend Data

Spend indexes from other platforms moved far less over similar stretches. On the Databricks AI Index, the median user’s weekly spend rose 6.2x in 15 weeks to the week of 2 November, while spend at the 99th percentile rose about 1.4x (our piece). On the Ramp AI Index, the top 1% of US firms paying for AI spent a median $10,199.20 per employee per month in October, 3.16 times the January level (our piece).

These are different measures. OpenRouter counts tokens on its own API traffic, Databricks counts dollars on its gateway and Ramp counts dollars on corporate cards, so the multiples are not directly comparable. Companies are already setting limits: Bloomberg reported on 4 November that Apple cut some engineers’ daily AI budget to $200 from $300 (our piece).

The Structural Read

As we read OpenRouter’s two posts, the 106x is mostly a story about re-sent context: an agent sends its whole history on every step, and cached prompt tokens were just under 90% of agentic tokens in September, per the newsletter.

As we read the classifier, agent and human are behaviours of an API key, not of a person: tool calls, turns, request timing and cache use decide the label each month.

As we read the spend data next to it, token counts and dollar budgets are moving at very different speeds, 106x against 6.2x and 3.16 times on the Databricks and Ramp indexes, which measure different things.

OpenRouter, Inferred newsletter, 8 October 2026

“In the last week of September, agentic API keys used about 7x the tokens from human API keys”

Three Implications

FOR DEVELOPERS OpenRouter’s newsletter says cached prompt tokens were just under 90% of agentic tokens in September, so the cache rate weighs heavily on what an agent costs to run.

FOR MODEL PROVIDERS Agentic keys took 76% of OpenRouter’s tokens in October, about 600 trillion, against 8% a year earlier, OpenRouter said.

FOR FINANCE TEAMS Bloomberg reported that Apple cut some engineers’ daily AI budget to $200 from $300, one of the limits companies are already setting.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework says: “As companies move from chatbots to thousands of autonomous agentic workflows running 24/7, the line item that used to be a rounding error becomes the budget.”

As we read it, OpenRouter’s chart shows that shift on one platform’s traffic: agent tokens up 106x this year against 7x for human tokens. The framework does not discuss OpenRouter or this chart.

What Is Not Established

The image carries a source line and no method note, so we take the agent and human definitions from OpenRouter’s own recap and newsletter; the chart does not say where keys classed as mixed fall. Its title says percent growth while its axis and labels are in multiples (106x, 7x), and the post gives no weekly token counts. The figures cover traffic through OpenRouter only, not the wider AI market.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The CFO’s Guide to the Token Economy

A Business Engineer framework on how token spend becomes a budget line as companies move to agentic workflows.

Read the Map of AI →

The Bottom Line

OpenRouter’s data, as charted by its head of insights, puts weekly agent token usage up 106x since the first week of January against 7x for human usage, about 15 times the multiple by our arithmetic. By OpenRouter’s own figures, agentic keys used about 7x the tokens of human keys in the last week of September and took 76% of October’s 790 trillion tokens.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. We read Peter Walker’s post on X of 16 November 2026 and the chart attached to it, OpenRouter’s Inferred newsletter of 8 October 2026 and OpenRouter’s October 2026 data recap in full; the token multiples, definitions and shares are OpenRouter’s. Walker’s role is from his X bio. Databricks, Ramp and Bloomberg figures are from our earlier pieces, linked above. The ratio of the two multiples is our arithmetic.

Sources: Peter Walker (OpenRouter) on X (16 Nov 2026) · OpenRouter Inferred: Cache is Crucial (8 Oct 2026) · OpenRouter: October 2026 data recap (6 Nov 2026)

Scroll to Top
FourWeekMBA