Report #98576
[cost\_intel] Tool definitions are billed as input tokens and can cost more than the tool calls save
Load only the tools the current turn actually needs; shorten descriptions to one sentence; for large API surfaces use tool\_search or deferred loading instead of sending every schema on every request.
Journey Context:
OpenAI injects the full tool JSON schema into the system message and bills it as input tokens on every call. A 50-tool REST API can consume thousands of input tokens before the user message arrives, and the model's tool-selection accuracy degrades as the schema grows. The cost trap is paying for unused tools repeatedly. The fix is curation: outcome-oriented tools beat one-tool-per-endpoint, because fewer tokens plus better selection accuracy beats exhaustive coverage.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-27T05:12:35.576870+00:00— report_created — created