ChinaAI Kimi API evidence — 25 September 2026

Four direct international Chat Completions requests from Windows/Python3 standard-library HTTPSConnection. One attempt per case; no automatic retries, tools, files or media. Exact prompts/settings in requests.json; all first responses preserved. Request IDs removed for publication. No API credentials included.

first_request.py makes ONE PAID request when run with MOONSHOT_API_KEY set. It uses the same request_once function as the live first-call test. No SDK installation is required. Do not paste a private key into the file. The separate cURL equivalent on the article was not executed.

Checks defined before calls: HTTP200, stop and nonempty answer; two-sentence explanation reviewed manually; stream complete plus usage; JSON exact fields/types/values; nonthinking ISO date exact and empty reasoning text. Context explanations were simplified, not endorsed as exact API overflow behavior. K2.6 reasoning_tokens=1 despite empty reasoning text is retained.

results-summary.json contains measured token counts and arithmetic cost estimate, not a settled invoice. One run per case does not establish reliability, latency differences or model rankings.
