Claude Sonnet 5

Anthropic's active Sonnet model for general reasoning, coding, and agentic tasks. These notes describe model behavior, not Claude Code behavior.

• Give clear, direct instructions and specify the desired output and constraints.
• Explain relevant context and motivation rather than expecting unstated conventions.
• State the scope of each instruction explicitly; do not rely on the model to generalize it from one item to another.
• Choose effort deliberately: lower levels suit short or latency-sensitive work, while high or xhigh effort is more appropriate for difficult coding and agentic tasks.
• For complex prompts, separate instructions, context, inputs, and examples with consistent XML tags; put long documents before the query and ask for relevant evidence when grounding matters.
• Use positive examples and explicit verbosity or tone requirements when output shape matters.

• Relying on vague requests when completeness or output structure matters.
• Applying behavior observed in another Claude generation without re-evaluation.
• Assuming omitted scope will be inferred, especially at low or medium effort.

Adaptive thinking is enabled by default on Claude Sonnet 5. In API integrations, manual extended-thinking configuration is unsupported, non-default sampling parameters are rejected, and the output budget must leave room for thinking and tool calls. Model behavior is nondeterministic; validate guidance with representative evaluations and revisit it when the model-specific documentation changes.