Model routing
The practice of a product silently choosing which underlying model to run your query on — often based on cost, load, or query type.
Model routing is when an AI product routes different queries to different underlying models, often without telling you. A simple query might go to a fast, cheap model; a complex one to a larger, slower model. The user experience is one interface; the actual model behind each response varies.
Some products are explicit about routing ("we automatically pick the best model for your question"). Others are silent. Silent routing is controversial because you can't tell why one answer is worse than another — you might be paying for premium access but getting served the cheap model for cost reasons the product doesn't disclose.
Legitimate reasons for routing: - **Cost efficiency**: not every query needs the biggest model. - **Latency**: smaller models respond faster; some queries prioritize speed. - **Specialization**: a code-specific model may outperform a general model on code tasks.
Problematic patterns: - Silent downgrades under load (you paid for premium; you got the cheap tier because the provider was busy). - Non-disclosure of which model actually answered (you can't reason about consistency across your work). - Routing based on account age or tier without telling you (fairness concerns).
The transparent alternative: pick one good model, use it for everything, be honest about it. This costs more per query but earns trust.