Agentic AI now burns more tokens than humans: Cost implications for GTM

The Gist

• OpenRouter processes 28T tokens weekly, 1% of global inference
• Agent usage now exceeds human usage in token consumption
• Single agent task can cost 100x human chat due to context loading
• Most companies still budget AI costs like simple chat interfaces

• Agentic AI now consumes more tokens than human usage, significantly impacting cost structures for businesses.
• Agentic tasks require heavy context loads, leading to token bills that can dwarf human chat interactions.
• Inference quality and tool-call success rates vary by provider, even for the same model, affecting agent performance.
• Tool calling has become central to agentic AI, with 55% of requests asking for tools and 46% of completions finishing due to tool calls.
• Companies must budget for agentic AI as a multiple of human usage, not an extension, due to higher token consumption.
• Routing and failover layers are now core infrastructure for reliable agentic AI performance.

• Forecast AI spend based on agentic usage (a multiple of human chat costs), not legacy chat-based models.
• Evaluate and monitor tool-call success rates by provider, as this critically impacts agent performance.
• Treat routing and failover infrastructure as core to AI architecture, not optional.
• Benchmark inference quality across providers, even for the same model, to ensure consistent performance.

• 28 trillion tokens processed in a single week (OpenRouter's expected weekly token processing volume, representing ~1% of global inference.)
• 55% of requests ask for tools (Percentage of requests for a frontier model family on OpenRouter that involve tool calls.)
• 83% tool usage rate (Percentage of times the model uses tools when requested.)
• 46% of completions finish due to tool calls (Percentage of agentic tasks completed because of tool calls.)
• 213 tool calls (Number of tool calls tested in a live demo, showing provider-dependent success rates.)