Anthropic and OpenAI have both admitted in writing that their models got worse after launch, so I dug through the postmortems, the daily trackers, and the launch benchmarks to find out what's actually changing. I couldn't find proof that the model itself gets swapped, but there are four dials around it that they do turn down: effort, the instructions, routing, and the app itself. I also get into why compute pushes them to do it, and why nothing makes them tell you when they do.