routing strategy

A routing strategy means not handing every task to a single AI, but splitting the work among multiple AIs depending on what each task requires. It's similar to a hospital sending difficult diagnoses to an expensive specialist while routine checkups and prescriptions go to a general practitioner. The expensive, high-performing model is reserved for critical, direction-setting moments, while the cheaper, lighter model handles repetitive, simple work.

In coding, for example, you might ask a powerful, expensive model to work out the overall design or make tough judgment calls, then hand off the repetitive job of actually writing the code to a much cheaper model. This way, you don't have to use the most expensive option every single time, which cuts overall costs substantially without a major drop in output quality.

This kind of division is necessary because every model has a different balance of performance and price. The core of a routing strategy is the habit of allocating tasks appropriately, keeping in mind that the highest-performing model is always the most expensive, and that not every task actually needs top-tier performance.

In the article, a developer's comment that "I talk through the design direction with Fable, then hand the actual work instructions to GLM" is an example of a routing strategy in action. This isn't the name of a specific product feature — it refers to a user's own judgment and habit of splitting tasks based on cost and performance.

Try giving the same question or coding task to two AI models with different performance levels, then compare the results, time taken, and cost. For instance, ask a complex design question to the expensive model and a request to actually implement that design to the cheaper model, then check how much the quality of the outputs differs. This will give you a feel for how a routing strategy works.