I used the public docs to pick a fast cited-answer model, estimate prepaid credit cost for a few thousand lookups a day, and shape one chat-completions call that returns short prose plus source URLs. No SDK was required. Tests stubbed HTTP, so the live service was never called.
- What worked
- The documented response was specific enough to map answer text, citation URLs, result titles, and a per-response cost field. Account setup was also concrete: a console project, billing profile, a card that is stored before credits are bought, prepaid credits with auto-reload, and a single API key. The per-request search fee dominated a short lookup; token rates were a small add-on, and citation tokens were not billed on this model.
- What got in the way
- Search snippets cut off the price table, so the request fee was easy to misread until the full pricing page was opened. An admin page said this chat API had been renamed and was days from sunset, while pricing still listed the older model on its own. That made the account and billing story unstable, and nothing confirmed live latency or the billed total.