ChinaAI.chat — Kimi vs ChatGPT official-app test evidence
Research: 24 September 2026. App collection/checks: 24–25 September 2026.
Kimi: international kimi.ai, Moderato, K3 High. Fresh tasks; existing account configuration preserved. No custom skill invocation was observed; hidden personalization was not ruled out.
ChatGPT completed results: Pro, Chat, Temporary/Unpersonalized. Latest/Extra High visible for tasks1 and3; task2 inherited setting. Underlying checkpoint not exposed.
Eight completed requests: three initial tasks and one predefined HTML revision per app. One preliminary ChatGPT Free request failed; no retry replaced that failure. Work attachment/model mismatch blocked preparatory task2 steps; actual task2 ran in Chat.

PROVENANCE
Responses are extracted visible answer text; citation links are separate. The sole redaction is an internal Kimi preview-version ID. Some interface labels are retained. Model-generated claims about accessibility, privacy or self-testing are not independently verified merely because they appear in a response.
Both apps' two CSV outputs were actually downloaded and parsed. Kimi CSVs came from download-all ZIP; Pro CSVs came from direct downloads.
Kimi HTML v1 and v2 are verbatim original source recovered from the app's code viewer, NOT verified HTML downloads. ChatGPT HTML v1/v2 are original downloads. Original flaws are preserved; no repairs were applied. The hosted Kimi preview DOM is omitted because it contains separate preview instrumentation.
Screenshots show original tools running locally in an actual browser. They are not mocked up or redesigned. No private chat URL, account identity, credential or unrelated content is included.

RESULTS
Task1: both passed seven predefined factual checks on six draft claims; rewrites220/209 words respectively. Both simplified the supplementary USB-PD trigger to its power condition; full rule also includes voltage/current alternatives. Not a general accuracy ranking.
Task2: both match reference counts/medians and sole T05/E05-2 exception. Kimi explicitly warned about small sample size; Pro final prose did not restate it. Fictional inputs, not real support records.
Task3: both preserved exact six records and combined filters, shortlist totals/retention, reset and empty state after one revision. Kimi loses keyboard focus after an Enter card toggle. Pro retains that focus but loses it after removing a shortlisted item whose card is filtered out. Both need review before publication. Both showed no document-level overflow at390px; not a complete accessibility or cross-browser audit.

REPRODUCE
Submit identical prompts and original inputs in fresh task contexts. Keep the reference answer separate from apps. Record exact account/route/model labels, all failures and interventions. Download and inspect actual files. For HTML, open the retained local file and shortlist W1/W4 (165 minutes); apply Craft+available+<=90 (no matches but shortlist retained), then reset (six results, shortlist retained). Test keyboard controls separately. All workshops are fictional; shortlisting is not booking.
Do not treat these unequal paid-account runs as a free-tier comparison, equal-price trial, model benchmark, privacy audit or reliability estimate. No API calls or purchases were part of this work. No generated tool was deployed publicly.
