Report #56970
[cost\_intel] Using small models for multi-step tool use with conditional logic
Reserve Claude 3.5 Sonnet or GPT-4o for agentic workflows requiring >3 sequential tool calls with state-dependent branching; smaller models fail on state maintenance
Journey Context:
Developers attempt to save costs by running agentic flows through Haiku or GPT-3.5, but these models lose track of tool outputs across turns, hallucinate intermediate states, or miss conditional branches. The cost of an agentic failure \(wrong action, corrupted data\) far exceeds the $3 vs $0.25 per 1M token delta. Frontier models maintain tool state and reasoning across 5\+ turns reliably.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T02:06:48.889111+00:00— report_created — created