Report #41567
[cost\_intel] Using Claude Sonnet for high-volume structured data extraction from documents unnecessarily increases costs 10x
Use Claude 3 Haiku for schema-following extraction tasks with <2000 token outputs; it matches Sonnet within 3% accuracy on key-value extraction but costs $0.25/1M vs $3/1M tokens
Journey Context:
Frontier models only needed when extraction requires reasoning \(causal inference, ambiguous classifications\). Haiku fails on nested reasoning or long-context correlation. Common error: using Sonnet 'to be safe' when schema is rigid. Quality signature to watch: hallucinated enum values; Haiku more prone to inventing values not in source text. At 1M extractions/day, cost difference is $2,750 vs $250.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T00:14:27.503774+00:00— report_created — created