Costs Aren't Coming Down. So What Now?

According to Goldman Sachs, token usage is projected to grow roughly 24x by 2030. As a result, AI leaders' concerns are shifting from "agent performance" to "whether costs can be sustained at scale."

Companies are currently facing two major challenges in optimizing their agents.

• Limits of manual tuning: New optimization methods emerge every week, and manually setting routing rules can actually worsen performance and cost as models or workflows change.
• Impact of model updates: Every time the underlying model changes, existing pricing and performance assumptions become obsolete, resetting optimization efforts.

To solve these problems, the industry is focusing on routing technology. According to a Vercel report, teams operating at scale use an average of 35 models in their routing graph.

To automate this, AI21 is building an intelligent router. This layer detects and removes token leakage at execution time—such as duplicate context or unnecessary instructions—and, as traffic accumulates, it learns to make increasingly sophisticated routing decisions that forward requests to the optimal model.