Judge AI costs by useful results, not token prices
AI value should be measured by completed work and the full cost of reaching an acceptable result, not by token price alone. OpenAI says the price per million tokens fell 97% from GPT-4 to GPT-5.4, while GPT-5.6 used 54% fewer and took 57% less time per task in a coding test. A cheap model can still cost more overall if it fails, retries, or produces work that needs extensive correction.
Models should be tested on real tasks and difficult after defining what counts as good enough; the calculation should include tool use, attempts, completion rate, processing time, and . Spending should be tied to visible usage and directed toward repeatable workflows with clear ownership and measurable value. Clear instructions, limited tools, reusable context, and explicit stopping rules can prevent wasteful loops.
Workflows that access company data or outside systems need , privacy controls, approval steps, and before they grow. Smaller or faster models suit work where they meet the quality bar, while stronger models should be reserved for complex, uncertain, or high-risk tasks.
Key points
- Measure the full cost of one accepted result, not just the model's token price.
- Test models on real work and with a quality bar set in advance.
- Use clear instructions and stopping rules to reduce retries and wasted spending.
- Invest first in repeatable workflows whose value can be measured.
- Set access, privacy, and approval rules before connecting sensitive data or outside tools.