Report #93515
[cost\_intel] Using reasoning models for single-tool API calls or simple function calling \(1-2 tools\)
Reserve reasoning models for tool chains requiring >3 sequential tools with conditional branching; use function-calling-optimized instruct models \(GPT-4o/Claude-3.5-Sonnet\) for 1-2 tool operations; reasoning adds 15s latency per tool call
Journey Context:
Reasoning models 'think' about tool schemas unnecessarily; for simple GET requests, instruct models achieve 99.5% success rate at <500ms; reasoning models achieve 99.8% at 15-20s—not worth 50x cost increase; the breakpoint is conditional logic complexity: when tool B selection depends on tool A's result which depends on a logical condition, reasoning becomes necessary
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T15:33:08.485561+00:00— report_created — created