What Kimi K3 & GLM-5.2 Actually ChooseChina's AI models vs Claude Sonnet 5

Three coding agents answered the same 60 devtools questions twice: once in English, once in natural developer Simplified Chinese, the script of mainland China. The agents: Claude Sonnet 5 inside Claude Code, Moonshot's kimi-k3 inside Kimi Code CLI, and Z.ai's GLM-5.2 inside ZCode. Every question ran against its applicable repos (2 to 5 of the 5), in two full runs per agent.

The twelve categories were picked for one property: each has at least one serious China-domestic alternative to the Western default. Object storage has Aliyun OSS and Tencent COS against S3 and R2. Maps has Amap and Baidu against Google and Mapbox. LLM APIs have DeepSeek, Qwen, and the two models' own makers against OpenAI and Anthropic. If a model carries a preference for either stack, these categories give it every chance to show.

The prompts never name a vendor and never mention China or compliance (one hosting prompt says “expanding to Asia”; the rest are geography-free). The only intended difference between the two versions of a prompt is the language; run-to-run noise is measured separately below.

Claude converges with the Chinese models only partway, and where it stops is the informative part. Claude's Chinese answers move in the same direction as Kimi's and GLM's, about half as far: it names a Chinese vendor in 23.9% of Chinese answers against Kimi's 32.5% and GLM's 49.3%, and it branches its advice by market in 28.2% of Chinese answers against their ~58%.

Where the Western default physically breaks in China, Claude converges with the Chinese models almost exactly. Its Chinese SMS picks are 54% China-domestic (theirs: 55% and 56%). Its Chinese payments picks are 36% domestic (33%, 39%). The statements that FCM does not reach Chinese Android devices and that Chinese consumers pay with Alipay appear in all three models' answers.

Where the Western tool still works, Claude re-orders less. On Chinese maps prompts, Kimi and GLM promote Amap to the top pick (63% and 80%); Claude keeps Google Maps first (70%) and adds Amap beside it (53%). And in the LLM API category the gap becomes absolute: Kimi and GLM answer the Chinese question with DeepSeek at 80% and 72%, while Claude's Chinese-language LLM picks are 0% Chinese. It answers Anthropic, every time, in both languages.

Claude does not localize the way the Chinese models do. Its answers cite the same failures in China, it holds its Western defaults longer where those tools still work, and on the LLM API question it does not move at all.

A reasonable reader will ask whether the English-side convergence means the Chinese models trained on Claude's outputs. Our design cannot answer that, and we want to be precise about what the data supports.

What we observe: Kimi matches Claude's top pick in 62.7% of English prompts, GLM in 59.0%, and the two Chinese models match each other in 62.1%. The symmetry is the point. Kimi's agreement with Claude is no higher than its agreement with GLM. We found no Claude-specific recommendation signature in these tests. The pattern is compatible with several training, post-training, harness, and corpus explanations that this design cannot distinguish.

A sharper cut of the same data asks who pairs off when the three disagree. On the 290 English prompts all three models answered, all three pick the same tool 47% of the time. When exactly two agree against the third, the pair is Claude+Kimi in 16.2% of prompts, Kimi+GLM in 14.5%, and Claude+GLM in 11.7%. Claude+Kimi leads, but a spread that size across three pairs is within sampling noise at this n (chi-square p ≈ 0.35). In Chinese the pairing counts are Kimi+GLM 18.7%, Claude+Kimi 17.0%, and Claude+GLM 11.9% (p ≈ 0.10 before accounting for clustering, so we treat it as no finding). Neither language shows a lopsided Claude-pairing.

What the Chinese-side depth does and does not show: GLM surfaces second-tier domestic vendors (Sensors Data, Qiniu, JPush, Authing) that Claude names rarely or never, which is hard to square with the strongest claim, that these models are Claude end to end. It does not speak to their English side, since native Chinese data and Claude-derived English data could coexist in one model. The English-side evidence is the agreement symmetry and pairing analysis above and the divergence test below. Their answer styles also differ from Claude's in stable ways (double the market branching, lower run-to-run stability, zero English self-promotion against Claude's 100%), though style is the weakest of these signals: post-training could reshape it either way.

We then ran that stronger test. Our March 2026 Codex study gives per-prompt reference picks for Claude (Opus 4.6) and Codex (GPT-5.3) on a different 12-category suite; 48 repo-prompt pairs have stable reference picks (the same tool in at least 2 of 3 runs) that disagree between the two agents. We ran Kimi K3 and GLM-5.2 on all 48 (one run, July 2026), extracted their picks the same way, and scored which side each landed on.

Kimi matched the Claude-side pick 22 times, the Codex-side pick 18 times, and neither on the remaining 8 (55% Claude-side of 40 decided, binomial p = 0.64). GLM matched Claude 11 times and Codex 14 (44% Claude-side, p = 0.69), and picked something neither reference chose on 21 of its 46 scored prompts, nearly three times Kimi's rate. We did not detect a Claude-side preference in either model. The marker tools tell the same story from both directions: Bun, which March-Claude picked in 63% of runtime prompts and Codex in 2%, shows up in 3 to 4 of the Chinese models' 10 runtime answers; Statsig, which Codex picked 27% of the time and Claude never picked once, shows up in 17 to 24% of their feature-flag answers. Each Chinese model carries markers from both Western agents.

Caveats on this test: the reference picks are four months older than the new runs (model drift blurs the signal in both directions), and the new side is a single run. Neither model showed a statistically detectable Claude-side preference in this sample: the 95% intervals on the Claude-side share are roughly 40-69% for Kimi and 27-63% for GLM among decided cases. The estimates are imprecise, do not rule out smaller preferences, and do not identify training provenance. The chi-square figure tests only whether the three exclusive-pair frequencies are equal.