The X-high thinking tier of Qwen3.8 27B consumes an extremely large number of tokens (e.g., 11,000 for drawing a pelican and 35,000 for drawing a dragon), while the medium tier is the best cost-performance sweet spot.