Prompting Claude Sonnet 5
How to prompt Claude Sonnet 5, the fast, cost-efficient tier. Read the shared Prompting guides principles first; this guide carries only what is distinctive to Sonnet 5. It performs well out of the box on existing Sonnet 4.6 prompts — the steers below are the behaviours that most often need tuning.
When to reach for it
Sonnet 5 has particular strengths in coding and agentic tasks and is the tier for well-scoped work, high volume, or latency-sensitive workloads — where Opus 4.8 or Fable 5 would spend more than the task warrants. As a rough migration mapping: Sonnet 5 at medium is comparable to Sonnet 4.6 at high, and Sonnet 5 at high to Sonnet 4.6 at max.
Prompting principles
| Principle | Steer |
|---|---|
Effort defaults to high |
Same default as Sonnet 4.6. Raise to xhigh for the hardest coding and agentic tasks; drop to medium/low for cost- or latency-sensitive work. It respects effort strictly at the low end. |
| Adaptive thinking is on by default | Requests with no thinking field run with adaptive thinking — a change from 4.6, where they ran without. Pass thinking: {type: "disabled"} to turn it off. If it thinks too readily (common with large system prompts), steer it down; if it under-thinks at medium, raise effort. |
| Mind the tokenizer + budget | The new tokenizer produces ~30% more tokens for the same text, and max_tokens caps thinking plus response together. Budgets tuned for 4.6 may truncate — leave headroom at high+ effort or you may see stop_reason: "max_tokens". |
| More agentic than 4.6 | It reaches for tools and self-verification loops more readily. But with thinking disabled it is less likely to consider searching — if you rely on tool calls with thinking off, add an explicit nudge. |
| Be literal, state scope | It interprets prompts literally and will not generalise across items or infer unrequested work. To apply something broadly, say so. |
| Trust the built-in updates | It gives regular, well-calibrated progress updates on long traces. Remove "summarise every N tool calls" scaffolding; if miscalibrated, describe what you want with an example. |
| No sampling params | Setting temperature, top_p, or top_k to a non-default value returns a 400 error — new for Sonnet-class. Steer tone and variety via the system prompt instead. |
| Specify the design brief | Like Opus 4.8, it can settle into a default house style that reads wrong for dashboards, dev tools, fintech, or enterprise briefs. Give a concrete spec, or have it propose 3–4 distinct directions first — the recommended route to variety now that temperature is unavailable. |
Coding and review harnesses
- Interactive coding. Use
xhigh/higheffort, add an auto mode, and specify task, intent, and constraints fully in the first turn to maximise autonomy and token efficiency. - Code review. As with Opus 4.8, a harness tuned for an older model may show lower reported recall because Sonnet 5 follows "be conservative / high-severity only" instructions more faithfully. Tell the finding stage its job is coverage and filter separately, or set a concrete severity bar.