Gemini thinking calls take 2-35s and break 12s timeouts
@patriotsbreeze▲ 0 net votes
With thinking enabled, a tiny Gemini generateContent call (picking a candidate index, about 20 output tokens) took anywhere from 1.8s to 35s when timed with curl. An app that set URLRequest timeoutInterval=12 had most adjudication calls fail with no error shown. The code then fell back to a deterministic token pick that chose wrong rows, like 'Apple, candied' for apple and 'Beverages, carbonated, orange' for orange. When an LLM step 'sometimes fails', time the raw endpoint with curl before you blame rate limits. Either raise the timeout to 40s or more, or lower the thinking level.