Agents Are Now AI's Biggest Buyers. The Routing Layer Just Got a New Conflict.

On February 6, 2026, humans and AI agents consumed roughly equal numbers of tokens on OpenRouter. That was likely the last time. Since then, agentic usage has grown 14x, from 0.51 trillion to 7.3 trillion tokens. Human usage grew 2.8x over the same period. Agents are now the dominant buyer of AI inference.

The headline growth is real, with an important modifier. About 70% of agent token consumption comes from cached prompts, which are billed at reduced rates. The actual spend is rising, but more slowly than the raw token count implies. OpenRouter also skews toward open-weight models, which run cheaper than frontier models. Take the 14x as a signal of direction, not a proxy for revenue.

But direction matters here, because of what Stripe just bought.

Stripe agreed to acquire OpenRouter, which routes requests across 400-plus models from 80-plus providers. That routing step is not neutral. The same model can cost dramatically different amounts depending on which provider serves it: in June 2026, Llama 3.3 70B ranged from $0.10 to $1.04 per million input tokens across providers listed on OpenRouter, a 10x spread. The router picks the provider. The router is now owned by your billing provider.

That is a structural conflict worth planning for. Stripe's financial interest in routing efficiency will often align with yours. It may not always.

Our index tracks paid agent endpoints using third-party telemetry from Coinbase. Of 25,210 listed paid endpoints, only 3,782, or 15%, have more than one distinct paying buyer. Most listed endpoints are services calling themselves, not genuine two-party commerce. The token-volume story on OpenRouter is real, but the broader market of agents buying from unrelated services is still thin.

For builders running agents in production, three things worth doing now:

Log the actual model and provider your requests reach, not just what you requested. Routing can diverge from intent, and the cost difference across providers is material.

Quarterly: run identical requests directly to each provider OpenRouter might route you to and compare against what your bill shows. Routing should be consistently cheaper or faster than you could achieve yourself. If it stops being that, you should know before your costs tell you.

For workloads where cost predictability matters, pin the provider in your OpenRouter configuration rather than relying on dynamic routing. You give up some failover coverage. You gain a cost input that does not move without your consent.

Agents are the market now. The question is who controls the terms.