Meterwise guide / 21 November 2026
Choosing the right model for each task
Understand the price gap
The input price gap between Claude Fable 5.2 and Claude Haiku 5.5 is 80x. Output prices and cached-input prices have their own differences, and providers do not use identical model names or price structures. A workload's split between input and output therefore matters when comparing options.
The gap shows why model choice belongs in a cost audit, but it does not prove that every task can move to the cheapest model. Complex reasoning, nuanced writing and demanding extraction may need a different option from straightforward classification. Start with the requirement of the task and use price to identify candidates worth testing.
Route by task, not by habit
List the distinct tasks your application performs instead of treating all AI traffic as one workload. A customer conversation, a document summary and a metadata extraction step can have different quality thresholds and failure costs. Each may justify a different model even when they sit inside the same product.
Routing means sending each task to an appropriate option under explicit rules. Keep the routing logic visible and record which model handled each request in usage data. Review routes when prompts, models or task requirements change. Avoid a permanent default to the largest model simply because it was chosen during early development.
Test on your own requests
Build a representative sample from real requests, with sensitive material handled appropriately. Include ordinary cases and the difficult cases that matter to the business. Run the current and candidate models on the same inputs, then compare their outputs against criteria defined before you see the price saving.
Use human review or a reliable task-specific evaluation where possible. Check accuracy, completeness, format and failure behaviour rather than relying on a general impression of quality. Do not move to a smaller model on price alone. If the candidate is not good enough, test a different prompt or model and keep the existing route until the evidence supports a change.
Measure cost per unit of work
The total bill mixes price changes with changes in traffic. A growing bill can reflect more useful work, while a falling bill can hide falling usage. Measure cost per task, conversation or document so the cost of completing the work stays visible independently of volume.
Include input, cached input and output tokens in the calculation. Also watch retries and repeated attempts: a cheaper model can lose its advantage if it needs more attempts to reach an acceptable result. Pair the unit-cost measure with a quality measure and a volume measure. Together they show whether a routing change improved efficiency or simply shifted the work.
Introduce changes carefully
Begin with one well-defined task and a limited route. Compare observed unit cost and quality with the sample results, then expand only when production behaviour supports the decision. Keep a way to return traffic to the previous model if errors, retries or answer quality move in the wrong direction.
Make model choice an ongoing review rather than a one-time migration. New tasks, prompt edits and provider price changes can alter the best route. The pricing calculator helps estimate the token-cost effect, while your own evaluation supplies the quality evidence. An audit combines both into a ranked implementation plan for your actual usage.
Document why each route exists and who owns its review. That makes it easier to investigate a later change in cost or quality without repeating the original analysis. Keep the evaluation sample and production measures connected to the route, so the next decision starts from evidence about your workload rather than a general model comparison.