OpenAI’s GPT-5.6 family ships with roughly five to six reasoning-effort settings per model size, extending a now-standard feature that lets one model trade latency and cost for stronger step-by-step performance.
System-prompt conditioning appears to be the main control mechanism: labels such as low, medium, high—or Inkling’s continuous 0.0-to-1.0 scale—tell the model how much reasoning to spend, typically by changing token-length penalties during post-training or RL.
Benchmark evidence from GPT-5.6, gpt-oss and Inkling shows higher effort usually means more generated tokens and better scores, but gains taper at the top end, making maximum settings increasingly uneconomical.
Open-weight reports from DeepSeek V4, Nemotron 3 Ultra, Kimi K2.5, GLM-5 and Qwen3 suggest vendors mix several recipes—effort-conditioned SFT, RL with budget-aware rewards, and hard truncation budgets—to create on/off or multi-level reasoning modes.
The broader takeaway is that reasoning effort has become a second scaling knob alongside model size: smaller models at higher effort can sometimes match larger models at lower effort, though automatic mode selection remains unresolved.