How much of that is because the models are optimizing specifically for SWE bench?