Stop counting tokens, start counting what ships
Why rationing agentic coding backfires, and what to measure instead.
TL;DR
- AI coding agents require more tokens than basic prompts due to planning, testing, and error correction cycles.
- Per-token pricing makes valuable AI workflows expensive, leading companies to consider rationing usage.
- A better metric than token cost is cost per merged pull request, which measures actual software output.
- Optimizing AI spending involves using frontier models for complex tasks and open-weight models for routine work.
- Owned AI inference capacity can lower costs for high, sustained usage and provide greater control over sensitive code.
- Hybrid AI models, using both APIs and owned capacity, offer a balanced approach for enterprises.
- Focus on building an operating model where increased AI use improves economics, rather than rationing adoption.