How models learn to think harder, or less

The low, medium, high reasoning dial that now ships with models like GPT-5.6 looks like a simple prompt setting, but Sebastian Raschka explains that the behavior behind it is trained in, not bolted on. He groups the methods into three families. The simplest puts the effort level straight into the system prompt, so the model reads "Reasoning effort: high" and has been taught what that means. A second approach uses supervised fine-tuning, mixing examples that show the full thinking trace with examples that skip it, so the model learns both habits before reinforcement learning begins. The third shapes behavior during reinforcement learning by charging a different per-token penalty for each mode: a light penalty lets a high-effort run ramble toward the right answer, while a heavier one forces the low-effort mode to stay short.

The concrete numbers make the tradeoff clear. Kimi K2.5's toggle method cut tokens by 25 to 30 percent while holding benchmark scores steady, which is a direct saving on cost and latency. Thinking Machines' Inkling exposes a continuous knob from 0.2 to 0.99 rather than three discrete steps, though Raschka notes the gains flatten out at the top of the range. Nemotron 3 Ultra spends only about 2.5 percent of its reinforcement learning prompts on the medium setting, a reminder that these modes need deliberate training coverage. His read is that no single recipe wins, most teams combine fine-tuning for the behavior with reinforcement learning to sharpen it, and the next step is likely models that pick the effort themselves.

Why it matters

If you pay per token or care about latency, the reasoning dial is a real lever, not a gimmick: the same model can cost a third less on easy work when you turn it down. Raschka's breakdown helps you reason about why a given model's low mode may still be slow, and what to test before you trust the setting on your own tasks.