Controlling Reasoning Effort in LLMs

Who Should Read

An LLM API developer who wants to understand how reasoning effort settings in models like GPT-5.6 and gpt-oss affect cost and performance

Core Mechanics

• Reasoning models output intermediate steps
  • It doesn't mean they reason like humans; they write a step-by-step reasoning trace.
  • DeepSeek-R1 trained this with RLVR (reinforcement learning with 0/1 rewards on verifiable math and code problems).
• The <think> tags are cosmetic
  • They mark where reasoning starts and ends so a UI can separate or hide it.
  • They do not give the model any reasoning ability.
• Qwen3 has a reasoning on/off switch
  • enable_thinking=False is a hard switch that prefills an empty <think></think> in the response.
  • /think and /no_think are soft switches learned through SFT (supervised fine-tuning on example answers).
• gpt-oss picks effort via the system prompt
  • The system message contains "Reasoning effort: low/medium/high".
  • The model card figure shows higher effort raises both response length and accuracy.
• GPT-5.6's top settings show diminishing returns

Community Reactions

• A trick to extend reasoning by hijacking the token stream
  • One commenter says you can find the token marking the end of reasoning and replace it with "wait, but" tokens to force the model to keep going.
• Opinion that llama.cpp already offers a token reasoning budget
  • A commenter says it works with a custom message.
  • They also say the ideal is to start with a small budget and expand by complexity.
• Opinion that reasoning formalizes the "think step by step" prompt
  • A commenter notes that prompt trick was found two years after GPT-3 was released.

How to Apply

• Run the same prompts at each effort level and compare tokens and accuracy
  • Take 10 representative prompts, call them at low/medium/high, and record output tokens and correctness in a table.
  • Set the default at the point where accuracy gains flatten.
• Route effort by task difficulty
  • Use low for simple classification or summaries and high for code edits or multi-step reasoning.
  • Allow the top setting only when needed since its returns diminish.
• Cap reasoning with a budget on local models
  • If you use llama.cpp, set a token reasoning budget and start small.
  • Define a complexity criterion and raise the budget step by step.

Try It

# Paste into Claude or ChatGPT (simplified version of the article's method: mimic effort via system prompt)

Reasoning effort: {{low, medium, or high}}

Answer the question below. For low, give only a short conclusion without reasoning; for medium, show key steps only; for high, solve step by step and verify.
Question: {{question}}

Next step: run the same question at all three effort levels and compare answer length and accuracy.

Terminology

Related Papers

• Mercury 3: Diffusion at scale

  Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality

• Timestral: Mistral's time-series foundation models

  A 209M time-series model ranks first on three benchmarks and runs on CPU.

• DeepSeek V4.1 Flash 1017 (Max)

  DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.

• If you don't run your own LLM proxy, how do you know whats going on?

  Running your own LLM proxy shows which keys, models, and cache misses drive your usage.

• DeepSeek: DeepSeek v4.1 Flash 1017

  DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents

• DeepSeek-v4.1-Flash-1017

  DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length

• Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware

Related Resources

• https://magazine.sebastianraschka.com/p/controlling-reasoning-effort-in-llms
• https://sebastianraschka.com/books/#build-a-reasoning-model-from-scratch