Why Claude Haiku, not a bigger model

The useful model is not the biggest model available. It is the cheapest model that clears the accuracy bar for the task.

For a posting run, Claude Haiku stayed under $1 per 100–150 items while clearing the bar needed for classification. That makes it the practical default: spend more model budget only when the task proves it needs more reasoning.

Start with the bar

The evaluation question comes before the model choice:

1. Define the classification that has to be correct.
2. Test a representative set of messy inputs.
3. Keep the cheapest model that clears that bar.
4. Re-run the set when the prompt or input shape changes.

That sequence keeps “use a bigger model” from becoming a substitute for measuring the prompt and the data.

Cost is part of the design

The cheaper model is not automatically better. It is better when its accuracy is sufficient and its cost leaves room to run the workflow repeatedly. A model choice should survive both checks: the output is useful, and the system can afford to keep evaluating it.