A new research paper says enterprises could shave millions off their AI coding agent bills just by rerouting prompts to cheaper models when it's safe to do so.
Researchers built a router called Jev that classifies every prompt sent to AI coding agents, like Anthropic's Claude Code or OpenAI's Codex, then decides which model should answer based on a price sheet and how prompt caching works. The router only shifts work to a different model at points where no running conversation has to rebuild its cache - at the start of a session, in side lanes, or when a subagent launches. Testing the approach on an emulated 10,000-seat enterprise, using behavior patterns drawn from roughly 10,000 real sessions in public datasets, the researchers found the router could recover 14 to 21% of model spend at Anthropic's list prices as of September 21, 2026 - $3.3 million to $5.0 million a year for that size of deployment.
Most companies don't build these agent harnesses themselves; they buy them from vendors and inherit whatever model-selection defaults come baked in. That quietly hands the vendor's harness, not the enterprise, control over the effective price paid per session. The paper's sharper finding is a "crossover" effect: on long, tool-heavy coding sessions, the priciest model can cost less than a mid-tier one, because switching models mid-task means losing the prompt cache and paying to rebuild it.
This is a simulation, not a shipped product, and real enterprise AI bills rarely match tidy percentage estimates - but it's a useful reminder that in agentic coding, the router deciding which model answers matters almost as much as the model itself.