I think it also helps that speed and latency aren't as vital for coding and we can afford to let the model think for longer.