The Routing Fault Line: When GPT-5.6 Silently Becomes a Mini Model

The API response was the first anomaly. A user selects "GPT-5.6 Sol's Thinking" on the front end, pays the premium, and expects the full weight of the flagship model. The server logs, however, tell a different story: model: gpt-5-5-mini. Not a near-identical sibling. A smaller, cheaper, distilled version. For roughly 3% of requests, the system executed a silent downgrade. The user saw one thing. The code did another. This is not a trivial UI bug. This is a fracture in the production routing layer, and it exposes the hidden dependency between cost optimization and user trust. Tracing the invariant where the logic fractures, we find that the abstraction between the front-end promise and the back-end execution leaks, and we must measure the loss.