Latest models have a much better sense of time and token budget, so if you instruct it for a quick answer or to not burn many tokens it will try to respect that