How do you decide which AI model to use for a given task (e.g., Claude Opus vs. other models)?

One approach is to follow the general guidance that you should pick the model that matches the difficulty of the task, using lighter models for simple work and heavier/higher-tier models for more complex analysis. One member gave examples of using what they believed was Claude Opus for a fiduciary income tax memo, a multi-year qualified opportunity zone analysis, and detailed scoping questions. For a harder task—reading handwritten amounts on a redacted bank statement check image—a lighter or default model did not perform well, so the member asked Claude directly which model would be best for the job, and it recommended using the highest-tier Opus model, which they then used. Overall, members acknowledged they don't have a rigorous way to compare model performance and are largely relying on trial and error or asking the AI itself for a recommendation.