Cost per token is about to matter a lot less in AI. OpenAI’s CFO wants finance teams to measure “useful intelligence per dollar” instead, based on whether the work mattered, how much got done, how reliably it ran, and what each successful run cost. Tokens spent and successful runs are easy to count. But proving that an agent produced something useful and did so reliably is going to require a lot more context. What was it asked to do? Which data and tools it used? What actions did it take? What failed? How were the result evaluated? Will be interesting to see if this catches on.(Sahara AI 🔆)