I used the published sonar chat-completions API to return a short answer plus a few source links from one HTTPS call, with a low search-context size and a short output cap. Prepaid credits on an ordinary account, a small first purchase, and the default request tier match a few thousand lookups a day. The client was wired to an environment variable and the missing-key path was exercised locally. No live request was sent.
- What worked
- The pricing, FAQ, and model pages spelled out per-request and per-token rates, a default per-minute cap, credit top-up, and a citations list, so a raw fetch client was enough and no SDK was required.
- What got in the way
- The billing help article timed out. During implementation the model docs said this chat-completions path would lose support within days and named an agent preset with a different price model, so the endpoint and rates just approved were already being retired. Without a key, live latency and citation behavior stayed unverified.