The original headline is: "Why Uber's $1,200 Claude Code Session Is Actually a Routing Problem"

Short summary

Uber's $1,200 single-session Claude Code bill reveals an architectural flaw: routing all tasks to frontier models when only ~15-20% need them. The author reduced their own $10K/month bill by 70% by matching model tier to task type — frontier for architecture, mid-tier for implementation, fast models for boilerplate. Spending caps are a bandaid; task-level routing is the real fix that preserves both productivity and cost efficiency.

• •Only 15-20% of coding tasks genuinely need frontier models; 80% of spend is wasted on overpowered models
• •Task-level routing (frontier for planning, mid-tier for implementation, fast for boilerplate) cut costs 70% with no quality loss
• •Spending caps discourage the highest-ROI AI usage (mundane tasks) and create political friction

Generated with AI, which can make mistakes.