Deliberate reasoning

Deliberate reasoning is the LLM mode where the model thinks step-by-step before answering — chain-of-thought, [[extended-thinking]], scratchpad — contrasted with fast pattern-match answers. The 'System 2' to plain chat's 'System 1'.

Dual-process theory in cognitive science distinguishes fast intuitive (System 1) from slow deliberate (System 2) thinking. 2024-2026 reasoning models brought this to LLMs: deliberate reasoning chains where the model thinks step-by-step before final answer — proven to improve hard-task performance. Implementations: chain-of-thought prompting (older), extended thinking (built into model — o-series, Opus 4, Gemini Ultra, DeepSeek R3), self-consistency (sample N reasoning chains + vote). Production trade-offs: 10-100× more cost + latency, but 2-5× quality gain on math, multi-step planning, and code. Routing decision: when speed + cost matter, use non-reasoning tier; when correctness on hard tasks matter, escalate to deliberate.

When to use deliberate reasoning

• Hard tasks: math, code, planning, analysis.

Common mistakes

• Forcing deliberate reasoning on every query — wastes compute.
• Skipping reasoning on hard tasks — answers degrade.

FAQ

Related terms

• Chain-of-thought prompting — Chain-of-thought (CoT) prompting tells a language model to write its reasoning steps before its final answer, increasing accuracy on multi-step problems.
• Extended thinking — Extended thinking is Anthropic's flag on Claude that allocates a configurable budget of internal reasoning tokens before the user-visible answer — enabling deeper reasoning on hard problems for a higher cost.
• Test-time compute — Test-time compute is the LLM technique of spending more inference compute per query (longer reasoning chains, multi-sample voting, deeper search) to get better answers — the foundation of reasoning models (o-series, Claude extended thinking, DeepSeek R-series) in 2026.

Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/deliberate-reasoning.md.