Not tested: K2.8 · K3.1 Pro

Kimi K2.8 and K3.1 Pro are missing from this record because the free tier will not take the question. K3.1 Pro, a post-trained version of K3, was released in late October. K2.8 first appeared as a preview in Moonshot’s coding tools in September. For the past month, every attempt to put the carwash prompt to either model on a free Kimi account has returned the same message instead of an answer: “Too many people are chatting with Kimi; a subscription will grant you priority access.” A paid subscription clears it at once, and the earlier Kimi tests here were run on one.

It is not a one-off. A TechRadar editor could not get in at K3’s launch in July for the same reason. App-store reviewers report seeing the message for many months, and one says it appears before any free limit has been reached. Moonshot has not said whether the cause is capacity or a limit on free accounts. What the record can say is that the message disappears when you pay. No question was answered, so these attempts are not runs: they carry no score and count in no tally. For anyone not paying, Kimi’s newest models are currently an upgrade prompt.

Kimi K3

Kimi K3 (launched the week of July 13, 2026) reverses its predecessor’s record outright. Six days after Kimi K2.6 went one-for-eight across four languages in Carwash III, K3 and its parallel-subagent variant K3 Swarm went twelve-for-twelve — every tier (Standard / High / Max), every language (English, Chinese, French, Ukrainian), every run a Drive. The largest single-generation reversal in the dataset. Two textures: the Swarm orchestrator’s traces explicitly decline to spawn subagents for this question ("No need for skills, plans, or subagents"), and every non-English run reasons in English — recorded under the operator-readability standard (a trace the operator cannot read fails at making the reasoning inspectable), though scoring remains response-based.

Results

Transcripts