0
I’m evaluating Claude and OpenAI models for a tool that summarizes long reports. The headline price per token is easy to compare, but I’m not sure it tells the whole story when documents are large and the same instructions are sent with every request.
What should I measure in a small trial to estimate the real cost—input and output tokens, prompt caching, context limits, or how often the model needs a second pass? If you’ve compared them on a similar workload, I’d be interested in what changed the bill most.
