If the model is better in that it resolves the task with fewer iterations then the i/o token pricing may be a wash or lower.