AI Agents Are Already Outspending Humans on OpenRouter

OpenRouter routes requests across dozens of model providers, so its own usage numbers are one of the few places to see agent traffic against human traffic at scale rather than inside one vendor's walled garden. The numbers, reported by The Decoder in August, are stark. Since February 2025, token consumption driven by AI agents has grown roughly 14 times. Human-driven usage over the same period is up about 2.8 times. Agents now account for more tokens moving through the platform than people typing prompts by hand.

That gap is not a rounding error. It is the shape of where AI spend is actually going in 2026, and it changes what "AI budget" means for a team that is past the chatbot stage and into agent workflows.

Why agents burn tokens differently

A person chatting with a model sends one message, reads one answer, and maybe follows up. An agent working a task does none of that in a straight line. It plans a step, calls a tool, reads the result, decides whether to retry, and often re-sends a chunk of prior context with every one of those steps so the model has enough to reason with. A single task that would be one exchange for a human can be a dozen model calls for an agent, each one carrying its own token cost.

Multiply that by however many workflows a team has running unattended, and it is easy to see why agent traffic outgrew human traffic by a factor of five once both are compared over the same stretch of time.

The number that keeps the bill from matching the growth

Here is the part that gets lost when "14x" is the only number people repeat. The Decoder's reporting also notes that close to 70 percent of that agent token volume comes from cached prompts, not freshly generated ones. Providers price cached input tokens at a steep discount to fresh ones, because the model is reusing work it already did rather than reprocessing everything from scratch.

That cache share is the difference between a bill that grows 14 times and a bill that grows a fraction of that. Agent workloads are still not free, but the growth in actual spend runs well behind the growth in raw activity, the opposite of how usage-based pricing is usually assumed to behave.

What this means for budgeting an agent rollout

Three practical shifts follow from this, and none of them are about cutting agent use.

Budget per workflow run, not per seat. A support agent that resolves a ticket in six tool calls costs something specific and repeatable. That is a knowable unit cost the way a per-seat SaaS license never was, and it is the number worth tracking instead of a monthly total that just goes up.

Ask any agent framework or vendor directly whether it reuses context through prompt caching, and how. This is the lever that separates a 14x usage increase from a 14x bill, and it is not something every framework does by default. Frameworks that re-send full context on every step forfeit most of that saving.

Watch for retries and loops specifically, not total volume. A chatbot's cost is bounded by how much a human is willing to type. An agent's cost is bounded by how many times it is willing to retry a failing step before giving up, which is a setting, not a natural limit. That is where an agent workflow quietly turns expensive without anyone changing what it does.

The rule worth keeping

A rising agent token count is normal once a rollout is actually running, and OpenRouter's numbers show the market is well past the early-adopter stage here. The warning sign is a bill that grows in lockstep with usage instead of trailing it the way caching should make it trail. If those two lines are tracking together, the caching lever is not being pulled, and that is a configuration problem, not a cost of doing business.

For a team running more than one AI subscription or API alongside agent workloads, the same math applies at the stack level. AI Stack Optimizer runs your actual tasks and spend against current models to show which subscriptions are earning their keep before the next renewal, rather than after.