Claude 4.5
Claude 4.5 Sonnet / Haiku capabilities and usage tips
TL;DR
- Claude 4.5 is Anthropic's 2025 flagship model. Current main SKUs: Sonnet 4.5 (primary) and Haiku 4.5 (lightweight/low-cost).
- Good at: long document/table comprehension (200K default, optional 1M beta), structured output (JSON/tables), writing analysis, and stable code/tool calling.
- Model selection advice: start with Haiku 4.5 to get things working, then upgrade to Sonnet 4.5 for quality. Use 10-50 samples for eval -- check accuracy, format stability, and hallucination rate.
Key Focus Areas
When reading this page, pay attention to:
- Context window (200K default, enterprise can request 1M beta)
- Structured output capabilities (JSON/Tool Calling stability)
- Hallucination and factual question answering performance
- Code/table comprehension: improvements over Claude 3.5
For production use, lock down the output format (schema) and add self-check (e.g., require evidence/quoted snippets).
What's New (4.5 vs 3.x/3.5)
- Long context: 200K default; enterprise/specific endpoints can use 1M beta (requires dedicated Header).
- More stable structured output: JSON/schema constraints drop fewer fields, function calling is more robust in loops/error recovery.
- Code and tables: Better code comprehension/generation, more stable table/CSV extraction.
- Safety and alignment: Fewer refusals, but still recommended to output "source/confidence" for auditing.
Usage Tips
- Prompt structure: System message defines role/format/constraints, User provides context + task, add self-check when needed.
- JSON/tool calling: Temperature 0-0.3; require the model to self-check field completeness before output; limit tool calling rounds to avoid loops.
- Long documents: Summarize in sections then consolidate; for traceability, have the model output "quoted snippet + paragraph number."
- Latency/cost: Haiku 4.5 for high concurrency; Sonnet 4.5 for critical quality steps.