Prompting Claude
Claude now ranges from fast Haiku through balanced Sonnet and heavy Opus to Fable for the most demanding long-running agents. Choose the model and effort level together.
Haiku, Sonnet, Opus or Fable?
| Model | API input / output | Choose it when |
|---|---|---|
| Claude Fable 5.1 | $10 / $50 | you need maximum capability for long-running research, reasoning or agent work |
| Claude Opus 5.5 | $4 / $20 | you run long coding agents, computer use or complex knowledge work |
| Claude Sonnet 5 | $2 / $10 | you want the best balance for everyday code, analysis, writing, vision and tools |
| Claude Haiku 4.5 | $1 / $5 | low latency, high volume and inexpensive subtasks matter most |
Fable, Opus and Sonnet publish a one-million-token context window; Haiku publishes 200,000. Anthropic's general starting point for heavy workloads is Opus, but Sonnet has the same list price as GPT‑6 Sol and is a more economical starting point for many routine tasks. Test Haiku first on simple work. Escalate to Fable when Opus at high effort still does not pass.
Be explicit about the quality you want
Anthropic recommends describing the deliverable and constraints as if briefing a capable new colleague who does not know your local conventions.
Use XML when it clarifies structure
XML tags can separate instructions, context, documents, examples and variable input. The tags are not magic; the value is unambiguous structure and consistent names.
Good examples steer format and tone
Choose relevant, varied and structured examples. Anthropic suggests three to five when high consistency matters. Treat that as vendor guidance rather than a magic count: measure the effect and remove examples that do not cover a requirement or edge case.
For long context, put the material first
For very long document tasks, Anthropic recommends placing documents before the specific question, labelling sources clearly, and often extracting relevant quotations before synthesis. Still test retrieval and quality on your own document types.
Use effort rather than chain-of-thought formulas
Newer Claude models use adaptive or built-in thinking. Opus 5.5 always has adaptive thinking enabled and defaults to medium effort. Adjust effort against measured results. Ask for a short rationale, evidence or a verifiable intermediate deliverable when you need visibility.
A schema is not truth
Structured Outputs can enforce JSON and tool argument shape, but the application must still validate values, handle interrupted responses and verify facts. Prefilling the last assistant response is deprecated for current 4.6+ and 5-series models; use system instructions or structured output.
Sources
- Anthropic: Models overview — current family, context and price
- Anthropic: Choosing a model — official selection matrix
- Anthropic: Pricing — tokens, caching and tools
- Anthropic: Prompt engineering overview — success criteria and evaluation
- Anthropic: Prompting best practices — clarity, examples and long context
- Anthropic: Structured Outputs — schemas and limitations