Token economy analyzes the costs and trade-offs of language-model input and output consumption. Production systems expose the issue when an impressive prototype meets its first cost review: API spend becomes a unit-economics problem, while cheaper calls can erode quality if adopted without measurement. section 9.0
Use token economics to make optimization decisions with explicit cost and quality guardrails:
- Measure input and output separately; output tokens commonly cost three to five times more than input tokens, so reducing response verbosity can matter more than shrinking context. section 9.1.1
- Prefer focused context when it preserves task quality: sending less context can reduce hallucination surface area as well as cost. section 9.1.1
- Choose among frontier, workhorse, and cheap tiers by task requirements rather than treating the cheapest model as the universal optimization. section 9.1.1
- Check quality continuously after cost changes, because savings can create complaints and hidden quality erosion whose starting point is difficult to identify. section 9.0
A response can cost more than its prompt
At illustrative early-2026 rates, a 4,000-token input and 2,000-token output costs $0.012 plus $0.030 with Claude Sonnet, or $0.008 plus $0.016 with GPT-4.1. The shorter output still costs more than the longer input, making response verbosity an important optimization target. section 9.1.1
