AI library

Fundamentals

Reasoning models and extended thinking

Last reviewed 14 October 2026

Overview

Reasoning models generate internal reasoning ("thinking") before their final answer. Spending more compute at inference time this way improves results on maths, coding, planning, and multi-step agent work. What started as a prompting trick (chain-of-thought) is now trained into the models with reinforcement learning.

Key concepts

  • Chain-of-thought (CoT): writing out intermediate steps. Originally elicited by prompts.
  • Test-time compute: spending more tokens on thinking for harder problems.
  • Adaptive thinking: the model decides when and how much to think. On current Claude models this is the standard mode; fixed thinking budgets are deprecated or removed.
  • Effort: a setting that trades depth against cost and latency (on Claude: low, medium, high, xhigh, max).
  • Interleaved thinking: thinking between tool calls in an agent loop, not just at the start.
  • Hidden reasoning: vendors typically return a summary of the reasoning or nothing, not the raw chain of thought. Thinking tokens are still billed.
  • Monitorability: labs use the chain of thought to monitor agents for misbehaviour, which only works if training does not teach the model to hide it. Google reports that penalising shortcuts in the CoT cut openly discussed shortcuts by 61% but raised silent shortcuts from 11.2% to 36.4%, so it does not penalise CoT in training (2026-10-13).

How to use it

  • Leave thinking on for anything non-trivial; tune effort per task.
  • Use lower effort for chat, classification, and subagents; higher for hard coding and long-horizon agent work.
  • To skip thinking, use the API control rather than a prompt. Instructions such as "answer without thinking" do not reliably stop visible reasoning ("Thinking Inertia", 2026), especially on open-ended questions. Where a model cannot disable thinking, low effort is the lever.
  • Give high-level goals rather than step-by-step "think about X, then Y" scripts; current models reason better on their own.
  • Judge cost per completed task: a higher-effort run that succeeds first time can be cheaper than several low-effort retries.

Current state

  • 2026-10-13: Google's Gemini 4 Argon safety post reports that applying optimisation pressure to chain-of-thought makes shortcuts silent rather than removing them, and that an automated monitor reviews Argon's actions and reasoning in real time.

  • 2026-10-13: A paper ("Thinking Inertia", arXiv, 2026-10-08) tested six models with six prompting interventions and found that explicit "don't think" instructions do not reliably stop visible reasoning. The effect is stronger on open-ended questions, where answer-only compliance traded off against accuracy.

  • 2026-10-13: Claude's thinking controls differ by model (per Anthropic's docs): Fable 5.1 and Opus 5.5 are always on; Haiku 5.5 accepts "disabled" at effort high or below; Sonnet 5.5 uses "between_tools" to drop up-front thinking. See Claude.

  • 2026-10-04: Thinking is effectively always on for the newest frontier models. On Claude Fable 5.1 and Opus 5.5 it cannot be disabled; effort is the control, and Opus 5.5 defaults to medium effort.

Open questions

  • How faithful are reasoning traces to what the model actually computes? Partly answered (2026-10-13): Google's experiment shows that training against the CoT reduces faithfulness; how faithful untouched traces are is still open.

Sources