OpenAI, Anthropic Push Outcome-Based AI Cost Metrics

As companies move generative AI from pilots into production, headline prices per million tokens capture only part of the bill. Inference compute, tool calls, retries, latency and human review can make a cheaper model more expensive to operate. OpenAI and Anthropic are pushing the market toward outcome-based measures, including cost per successful task, so businesses can compare models against real workloads and manage dollar and token budgets more effectively.

OpenAI outlined its “useful intelligence per dollar” framework in July 2026, calling for total model, tool, retry and labor costs to be divided by the number of tasks meeting a defined quality threshold. It said prices per million tokens fell 97% from GPT-4 to GPT-5.4. On the Artificial Analysis Coding Agent Index, GPT-5.6 Sol scored 72.7%, versus 69.9% for Anthropic’s comparison model, with an estimated 36.2% lower API cost.