How to Cut AI Coding Costs 70% with Task-Level Model Routing

Short summary

Uber exhausted its annual Claude Code budget in 4 months by routing all tasks to expensive frontier models. By profiling work and routing—architecture to Opus, features to Sonnet, tests/formatting to Haiku—the author cut monthly spend from $10K to $3K while maintaining velocity. The strategy: task-level model selection prevents vendor lock-in and creates a structural cost advantage as AI workflows scale.

• •60% of coding tasks are mechanical work (formatting, boilerplate) that don't need frontier-tier models
• •Three-tier routing (frontier for architecture, mid-tier for features, cheap for tests/docs) reduces costs 70% without quality loss
• •Vendor diversity and task-level classification create competitive advantage as agentic workflows become standard

Generated with AI, which can make mistakes.