Report #80459
[cost\_intel] Silent 10x cost spike from JSON schema bloat in structured outputs
Minimize JSON schema keys/descriptions and use \`additionalProperties: false\`. For complex schemas, ask the model for Markdown/YAML and parse locally instead of forcing JSON mode.
Journey Context:
Structured Outputs forces the model to output every schema key, even if null. Output tokens are 3-5x more expensive than input tokens. A 100-line JSON schema with descriptions can bloat output tokens by 10x compared to a simple markdown list parsed by Pydantic. The degradation signature: massive output token counts full of '"key": null'.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T17:39:44.260975+00:00— report_created — created