AI agents may now be consuming more tokens than human users on OpenRouter, marking a shift in how demand for model inference is generated and paid for. The change matters because AI systems are increasingly producing work for other AI systems, rather than serving only direct human prompts.

According to reporting by The Decoder, citing OpenRouter analyst Peter Walker, February 6, 2026, may have been the last day on the platform when humans consumed more tokens than agents. Agentic token usage has risen 14 times since then, while human usage has increased 2.8 times.

The figures point to a rapidly expanding class of workloads in which one model calls another, continues operating for longer periods, or launches additional AI processes. They do not, however, mean that inference spending has increased by the same multiple. Much of the measured usage consists of repeated context that is served from cheaper cached prompts.

What OpenRouter’s figures show

The central data point is a comparison between token consumption by human-driven use and agent-driven use on OpenRouter. The Decoder reported that agents passed humans in total token consumption after February 6, based on Walker’s analysis. The report does not provide the underlying chart, absolute token totals, or a detailed definition of how OpenRouter classifies agentic traffic.

The reported growth rates are still significant. A 14x increase in agentic usage versus 2.8x growth in human usage suggests that autonomous or semi-autonomous workflows are expanding faster than conventional chatbot interactions on the platform.

The result should be read as a platform signal, not a complete measure of the AI market. OpenRouter’s traffic reportedly skews toward open-weight models, which The Decoder says are generally less token-efficient than models from OpenAI or Anthropic. Users of those models may therefore generate more tokens for comparable tasks, making the platform’s mix different from that of closed-model providers.

Token growth is not the same as cost growth

Nearly 70% of agent token consumption reportedly comes from cached prompts. Caching allows repeated instructions or context to be billed at substantially lower rates than newly processed input, so raw token totals can rise much faster than actual inference costs.

That distinction is important for product teams planning agent deployments. A workflow that appears expensive when measured only by tokens may have a manageable bill if most of its context is reused efficiently. Conversely, teams that treat caching as a guarantee could be exposed to higher costs when prompts change frequently, cache hits decline, or agents generate large amounts of new output.

The data also highlights a measurement problem. Token counts capture how much model context is processed, but not necessarily how much useful work is completed. An agent that repeatedly revisits the same instructions or calls several models before reaching an answer may drive substantial usage without delivering a proportional improvement in quality.

Why agents are driving demand

Agent workflows tend to consume more tokens than single-turn conversations because they preserve state, inspect intermediate results, call tools, and revise plans. They may also delegate subtasks to separate model instances. Each step can add context, tool output, reasoning traces, or instructions to the next request.

The Decoder connected the increase to agents working independently for longer stretches and spawning additional AI processes. That behavior is consistent with a broader move from chat interfaces toward systems that execute multi-step tasks, although the available evidence does not establish which specific applications or industries account for OpenRouter’s growth.

The report also places the trend alongside the rise of reasoning models. These systems can use more internal processing before returning an answer, increasing token consumption even when the user’s request is short. The Decoder characterized this as a source of “token inflation,” but the available evidence does not show how much of OpenRouter’s agent growth comes from reasoning models versus longer-running orchestration.

Implications for builders and enterprise buyers

For AI builders, the immediate lesson is that model selection alone will not determine operating costs. Agent architecture becomes equally important. Teams will need to monitor cache-hit rates, context size, tool-call frequency, retry behavior, and the number of model calls required per completed task.

Reliability is another concern. Longer-running agents create more opportunities for an incorrect assumption, failed tool call, or unnecessary loop to compound. A system that is inexpensive per request can still become costly when it runs without firm limits. Budgets, execution timeouts, approval gates, and clear stopping conditions are likely to matter as much as prompt quality.

Enterprise buyers should ask vendors for workflow-level metrics rather than headline model prices. Useful questions include how many calls an agent makes per task, what portion of traffic benefits from caching, how usage is measured across delegated agents, and whether customers can impose spending or execution limits.

The market implication is also notable. If AI increasingly generates demand for AI, model providers may be selling not only to human users but to software systems that continuously consume inference. That could support large volumes of usage, while also increasing pressure on providers to improve caching, orchestration, latency, and pricing transparency.

What to watch next

The first signal to watch is whether other major model platforms report a similar crossover between human and agent usage. OpenRouter’s open-weight-heavy traffic makes it useful evidence, but not a definitive market-wide benchmark.

Second, builders should look for more complete cost data. The 70% cached-prompt figure explains why token growth may overstate spending, but it does not reveal total revenue, average cost per task, or the share of uncached output. Those measures would show whether agents are materially expanding paid inference or mainly increasing low-cost context processing.

Third, adoption claims should be tested against task completion and retention. Rising tokens can reflect productive autonomous work, inefficient prompting, or both. Evidence about successful task rates, human intervention, and repeat usage will be more informative than volume alone.

Finally, model providers’ responses will matter. Improvements in context caching, smaller specialized models, and more efficient reasoning could reduce the cost of agent growth. At the same time, providers may introduce pricing and controls designed specifically for software-to-software demand.

Creati.ai perspective

OpenRouter’s reported numbers capture an important change in AI economics: the fastest-growing user of a model may be another model. But the evidence is still a platform-specific signal based on analyst commentary, not an independently verified census of agent activity across the industry.

For teams deploying AI agents, the practical priority is to measure useful work per dollar, not tokens alone. Agent growth can expand the addressable market for inference, but sustainable deployment will depend on controlling loops, proving task reliability, and understanding when caching genuinely lowers the cost of autonomous execution.