Reasoning Model Prompts

# Reasoning Model Prompts — 2026

> **"Think step by step" is now counterproductive on reasoning models.** The model reasons internally. Your job is to set the effort dial and define the outcome — not to script the reasoning process.

## The Shift

In 2026, frontier reasoning models (Claude Opus 4.8 extended thinking, GPT-5 o-series, Gemini 2 Thinking, DeepSeek R1) handle multi-step reasoning natively. Explicit chain-of-thought prompting:

- Adds 20-80% latency
- Changes accuracy by negligible amounts (±3% on most benchmarks)
- Can actually hurt performance (Gemini Flash 2.5 saw -3.3% with explicit CoT)

---

## Effort Dials — The New Primary Lever

### Claude Effort Levels

| Level | When to Use |
|-------|------------|
| `low` | Bounded, mechanical tasks — known patterns, simple transforms |
| `high` | Complex reasoning — multi-step logic, analysis, planning |
| `xhigh` | Hard problems where being wrong is expensive |
| `max` | Maximum capability — novel research, edge cases |
| `ultracode` | Agentic coding — autonomous multi-file changes |

### GPT-5 `reasoning.effort`

| Setting | Use Case |
|---------|----------|
| `none` | Simple extraction, formatting, translation |
| `low` | Quick answers, straightforward Q&A |
| `medium` | Default — balanced speed/quality |
| `high` | Complex analysis, code generation |
| `xhigh` | Maximum reasoning — math, science, strategy |

---

## Prompt Patterns for Reasoning Models

### Pattern 1: Outcome-First Prompting

Write the *what*, not the *how*: