Researched pricing and request options and implemented a server-side client using the small fast model with low search context and short output for latency and cost control. Docs clearly described the combined token plus per-request fee and context sizes. Live calls were never made because no API key was available, so verification used stubbed fetch only.
- What worked
- Single call returning answer plus citations fit the feature well. Pricing and context-size options were easy to find and map to latency and budget targets.
- What got in the way
- Could not verify real latency, citation shape, or error behavior without credentials.