Research
Edwin Ong & Alex Vikati · July 2026
What Kimi K3 & GLM-5.2 Actually Choose
(China's AI models vs Claude Sonnet 5)
We gave Sonnet 5, Kimi K3, and GLM-5.2 the same 60 devtools questions twice, once in English and once in Chinese, across 5 repos: 1,860 responses. Asked in English, the three models mostly agree. Asked the same questions in Chinese, the answers change.
12 categories · 5 repos · 2 runs each · 0 runtime failures
Claude Code / Sonnet 5 · Kimi Code CLI / kimi-k3 · ZCode CLI / GLM-5.2
TLDR
- •Ask in English and China's frontier models sound like Claude. Western tools take about 96% of their primary picks across these 12 categories, and Kimi K3's most common LLM-API recommendation is Claude itself.
- •Ask the same questions in Chinese and the domestic stack appears. China-domestic vendors jump from about 1% of Kimi K3's answers to 32.5%, and from 3% to 49.3% for GLM-5.2.
- •Every request came from a US IP address. The language of the question was the only thing that changed, and all three models, Claude included, started splitting their advice between users inside and outside China.
- •Claude Code was the only agent that picked its own maker every time: Anthropic in all 60 of its LLM-API answers, both languages. In English, Kimi Code and ZCode never picked theirs.
The three models
If you mostly follow Western AI, two of these names may be new. Both are frontier-tier releases from Beijing labs, both are weeks old, and both ship their own Claude Code-style coding agents.
Kimi K3
Moonshot AI, Beijing · released July 16, 2026
A 2.8-trillion-parameter model from the Alibaba-backed lab, nine days old when we ran it, announced as open-weight with the weights scheduled for release after our collection date. It topped Arena's Frontend Code leaderboard the week it launched, with reported wins over frontier US models on several coding benchmarks. API list price is $3 in / $15 out per million tokens, the same as Sonnet 5's standard list price (Sonnet ran a $2/$10 introductory price through August 2026).
GLM-5.2
Z.ai (Zhipu), Beijing · announced June 16, 2026
A coding-first model with MIT-licensed open weights and a 1M-token context window, served by roughly 30 hosting providers. Z.ai sells it through a flat monthly coding plan priced as the budget alternative to Claude. Its API runs about $1.40 in / $3 out per million tokens.
Claude Sonnet 5
Anthropic, San Francisco · the reference point
The mainstream Claude Code tier and the subject of our original picks study. It anchors the comparison: the same questions, repos, and extraction rules as the two challengers.
Release dates and pricing from the linked vendor pages, checked July 27, 2026.
Every category was chosen because a serious Chinese alternative exists: Aliyun OSS vs S3, Amap vs Google Maps, Alipay vs Stripe, DeepSeek vs OpenAI. The prompts never name a vendor and never mention China (one hosting prompt says “expanding to Asia”; every other prompt is geography-free). The Chinese versions are the same questions in natural developer Simplified Chinese, the script used in mainland China and Singapore. Every request went out from US IP addresses, and the intended experimental difference within each matched pair was the prompt language.
Repos Used
nextjs-saas
Next.js SaaS
python-api
FastAPI service
react-spa
React + Vite SPA
node-cli
Node/TS CLI
kotlin-android-app
Kotlin Android
In English, China's models give nearly the same answers as Claude
Share of primary picks that go to China-domestic vendors when the question is asked in English. All three models give the Western stack 95-96% of their picks.
Sonnet 5 (434 picks)
Kimi K3 (440 picks)
GLM-5.2 (449 picks)
Custom/DIY Chinese-founded global China-domestic
China-domestic share of primary picks, English prompts. Chinese-founded vendors with global adoption (TiDB, Agora, DeepSeek) are counted separately and add another ~2.5% for every model, Claude included.
Same top pick on the same prompt (English)
Sonnet 5 ↔ Kimi K3 (n=306)
Sonnet 5 ↔ GLM-5.2 (n=290)
Kimi K3 ↔ GLM-5.2 (n=293)
Share of prompts where both models named the same primary tool.
Kimi agrees with Claude and with GLM at essentially the same rate. The three models give close to the same English answers.
The English consensus runs deep: Vercel wins cloud hosting for all three models (50-56%), Stripe wins payments for all three (100%), PostHog wins analytics (57-87%), Twilio wins SMS (67-73%), Clerk wins auth (45-50%). On the English half of this suite, the Beijing models' top picks match Claude's on roughly six of ten prompts, and the category winners are the same list.
The clearest exception is a Chinese vendor winning in English. Two database prompts ask for a MySQL-compatible distributed database and for an HTAP database, and on both, all three models picked 
TiDB: 48 of 48 responses, in English and Chinese alike. TiDB is Chinese-founded (PingCAP), and it is GLM-5.2's top English database pick. Where a Chinese-founded product has global adoption, the language of the question stops mattering.
Translate the question, and the stack changes
Share of answers whose primary picks include at least one China-domestic vendor, by prompt language. Identical prompts and repos in both languages. Chinese answers also name more tools per response (2.1 against 1.5), so they have more chances to include a domestic vendor. The share-of-picks figures below are the length-controlled version of the same result.
Asked in English
“i want to add an ai chat feature to this app. which llm api should i use”
Recommendation: Anthropic's Claude API via the @anthropic-ai/sdk package. Claude models are strong at exactly what a deployment-CLI assistant would do — explaining commands, reading logs, diagnosing failed deploys.
Kimi K3 · node-cli · run 1
Asked in Chinese
“我想给这个应用加一个 AI 聊天功能，应该接哪家大模型 API”
推荐：DeepSeek 或 Kimi（Moonshot）… 国内可直接访问、可直接充值，不需要解决网络和境外支付问题… Claude / OpenAI：模型质量确实好，官方 TS SDK 也成熟，但国内直连和付费都麻烦。
Recommendation: DeepSeek or Kimi (Moonshot)… directly reachable and payable inside China, no network or foreign-payment hurdles… Claude / OpenAI: model quality is genuinely good and the official TypeScript SDKs are mature, but direct access and payment from China are a hassle.
Kimi K3 · node-cli · run 1
In Chinese, Western tools keep most of their share (63-76% of picks); the Chinese vendors get added next to them, usually inside an explicit “domestic users vs overseas users” branch. Kimi's Chinese answer above names DeepSeek and its own maker, then adds that Claude and OpenAI are “genuinely good” but hard to reach and pay for from inside China. That is a judgment about the user's situation, and the only evidence the model has for it is the language of the question.
The top pick itself changes between the English and Chinese versions of the same prompt in 60.9% of GLM-5.2's 289 pairs, 47.2% of Kimi's, and 30.4% of Claude's. Claude flips least, and its 30% translation flip still runs about 10 points above the 21% flip rate it shows between two identical runs. For Kimi and GLM the translation effect clears their own rerun noise by 18 and 26 points.
Claude Code recommended Anthropic in every LLM-API answer. In English, Kimi Code and ZCode never made their makers the top pick.
The LLM API category asks which model provider to build on: an AI chat feature, document summarization, a coding assistant. Each agent is answering a question about its own market. One caution before the chart: each model ran inside its maker's own coding agent, whose system prompts we cannot see, so these are claims about the shipped agent products, and Claude Code's answers read as already Claude-conditioned (its Chinese answer opens “if this feature is built with Claude”). A rerun through one neutral harness is the clean isolation test.
Rate of naming its own maker as a primary LLM-API pick
n=26 to 30 responses per cell; exact rates 100/100, 0/56.7, 0/48.3.
Asked in English
“i want to add an ai chat feature to this app. which llm api should i use”
For an AI chat feature, the Claude API is the right choice, and I'd steer you toward Claude Sonnet 5 as the default model.
Sonnet 5 · node-cli · run 1
Asked in Chinese
“我想给这个应用加一个 AI 聊天功能，应该接哪家大模型 API”
如果这个功能是用 Claude 来做，直接接 Anthropic 官方 API 是最直接的路径… 默认用 Claude Opus 5。
If this feature is built with Claude, going straight to the official Anthropic API is the most direct path… default to Claude Opus 5.
Sonnet 5 · node-cli · run 1
Claude Code named Anthropic as the primary pick in all 60 of its LLM-API answers, both languages, both runs. Kimi and GLM made their own makers a primary pick zero times in English. In English, Kimi's most frequent pick was Claude (57%), and GLM split between OpenAI (54%) and Gemini (46%).
In Chinese, both Chinese models do promote their makers, to roughly half of answers, but the pick they agree on is DeepSeek, a competitor to both: 80% of Kimi's Chinese LLM answers and 72% of GLM's. Each names its direct rival in about half its Chinese answers too. Claude answers the Chinese question the same way it answers the English one: Anthropic, 100% of the time.
Which categories change in Chinese, and which never do
China-domestic share of each model's Chinese-language picks, category by category. The gradient runs from SMS, where the Western default physically fails in China, down to search, where no model named a Chinese vendor in any language.
| Category | Model | Top pick, English | Top pick, Chinese | Chinese answers: CN-vendor share |
|---|---|---|---|---|
SMS & Push Chinese alternatives: Aliyun SMS, Tencent Cloud SMS, JPush | Sonnet 5 |  Twilio 73% | Twilio 79% | 54% |
| Kimi K3 | Twilio 67% |  Aliyun SMS 53% | 55% | |
| GLM-5.2 | Twilio 72% | Aliyun SMS 76% | 56% | |
Maps & Geo Chinese alternatives: Amap (Gaode), Baidu Maps, Tencent Maps | Sonnet 5 |  Mapbox 63% |  Google Maps 70% | 30% |
| Kimi K3 | Mapbox 40% |  Amap 63% | 29% | |
| GLM-5.2 | Mapbox 50% | Amap 80% | 42% | |
LLM APIs Chinese alternatives: DeepSeek, Moonshot Kimi, Zhipu GLM, Qwen | Sonnet 5 |  Anthropic 100% | Anthropic 100% | 0% |
| Kimi K3 | Anthropic 57% |  DeepSeek 80% | 36% | |
| GLM-5.2 |  OpenAI 54% | DeepSeek 72% | 35% | |
Payments Chinese alternatives: Alipay, WeChat Pay, UnionPay | Sonnet 5 |  Stripe 100% | Stripe 100% | 36% |
| Kimi K3 | Stripe 100% | Stripe 100% | 33% | |
| GLM-5.2 | Stripe 100% | Stripe 90% | 39% | |
Object Storage & CDN Chinese alternatives: Aliyun OSS, Tencent COS, Qiniu | Sonnet 5 |  Cloudflare R2 33% | Cloudflare R2 47% | 16% |
| Kimi K3 | Cloudflare R2 37% | Cloudflare R2 53% | 30% | |
| GLM-5.2 |  AWS S3 30% | Cloudflare R2 54% | 31% | |
Realtime & RTC Chinese alternatives: Tencent TRTC, ZEGO, Agora (CN-founded) | Sonnet 5 |  LiveKit / AWS IVS (4-way tie) 17% |  Agora 37% | 18% |
| Kimi K3 | LiveKit 20% | LiveKit 23% | 24% | |
| GLM-5.2 | LiveKit 26% | Agora 40% | 29% | |
Product Analytics Chinese alternatives: Sensors Data, GrowingIO, Umeng | Sonnet 5 |  PostHog 73% | PostHog 50% | 8% |
| Kimi K3 | PostHog 87% | PostHog 77% | 9% | |
| GLM-5.2 | PostHog 57% | PostHog 53% | 28% | |
Git Hosting, CI & Registries Chinese alternatives: Gitee, Coding.net, npmmirror | Sonnet 5 |  GitHub 40% | GitHub 25% | 23% |
| Kimi K3 | GitHub 40% | GitHub 28% | 24% | |
| GLM-5.2 | GitHub 40% | GitHub 40% | 26% | |
Auth & Identity Chinese alternatives: Authing, WeChat login flows | Sonnet 5 |  Clerk 50% | Clerk 40% | 11% |
| Kimi K3 | Clerk 45% | Clerk 55% | 25% | |
| GLM-5.2 | Clerk 47% | Clerk / Authing 25% | 22% | |
Managed & Distributed DBs Chinese alternatives: PolarDB, OceanBase, TDSQL, TiDB (CN-founded) | Sonnet 5 |  Neon 45% | Neon 55% | 5% |
| Kimi K3 | Neon 50% | TiDB 40% | 10% | |
| GLM-5.2 | TiDB 40% |  Supabase 55% | 19% | |
Cloud Hosting Chinese alternatives: Alibaba Cloud, Tencent Cloud, Huawei Cloud | Sonnet 5 |  Vercel 56% | Vercel 57% | 12% |
| Kimi K3 | Vercel 50% | Vercel 53% | 7% | |
| GLM-5.2 | Vercel 56% | Vercel 54% | 14% | |
Search Chinese alternatives: Aliyun OpenSearch (none surfaced) | Sonnet 5 |  Meilisearch 40% | Meilisearch 40% | 0% |
| Kimi K3 |  PostgreSQL FTS 50% | PostgreSQL FTS 45% | 0% | |
| GLM-5.2 | Meilisearch 47% | Meilisearch 60% | 0% |
Bold Chinese-column picks changed from the English top pick. Percentages are share of responses naming the tool as the top pick.
The Chinese vendors, briefly
For readers who know the Western column but not the Chinese one.
Alibaba Cloud (Aliyun): Alibaba's cloud arm, China's largest. OSS is its S3; Aliyun SMS its Twilio.
Tencent Cloud: The cloud of WeChat's parent. COS storage, SMS, TRTC video calls.
Amap (Gaode): Alibaba-owned maps, the default navigation app in China.
Baidu Maps: Baidu's maps, Amap's main rival.
Alipay: Ant Group's payment app, over a billion users.
WeChat Pay: Payments built into WeChat.
DeepSeek: Hangzhou lab famous for cheap frontier-class open models.
Moonshot / Zhipu: The makers of Kimi K3 and GLM-5.2, the models in this study.
JPush / Getui: Push-notification services that reach devices FCM can't.
Sensors Data / GrowingIO / Umeng: Product-analytics vendors; Umeng is Alibaba-owned.
Authing: Auth platform, the Auth0 analog.
Gitee: Code hosting, the GitHub analog.
npmmirror: Alibaba-run npm registry mirror.
Qiniu / Upyun: Independent object-storage and CDN providers.
Agora / ZEGO: Real-time voice and video vendors; Agora is Chinese-founded and NASDAQ-listed.
TiDB / PolarDB / OceanBase: Databases from PingCAP, Alibaba, and Ant; TiDB has real global adoption.
Asked in English
“i need to show a map with store locations in this app. which maps api?”
For a Jetpack Compose store-locations map, go with Google Maps + maps-compose.
GLM-5.2 · kotlin-android-app · run 1
Asked in Chinese
“应用里要显示门店位置的地图，用哪家地图 API？”
国内市场（主流推荐）：高德地图 ⭐ 通用首选… 腾讯地图 微信小程序/微信生态内… 国内不能直接用 Google Maps。
Domestic market (mainstream recommendation): Amap ⭐ the general first choice… Tencent Maps for WeChat mini-programs… Google Maps can't be used directly inside China.
GLM-5.2 · kotlin-android-app · run 1
The largest shifts are in SMS (all three models put 54-56% of their Chinese picks on domestic vendors), maps (29-42%), and payments (33-39%), consistent with access and payment constraints the models themselves describe. GLM's Chinese maps answer states “Google Maps can't be used directly inside China,” and its payments answers assume Alipay and WeChat Pay.
The categories that barely flip are just as telling. Vercel keeps winning cloud hosting in Chinese for every model. GitHub keeps winning code hosting. Clerk keeps winning auth for Claude and Kimi (GLM ties it with Authing and Supabase Auth at 25% each). And no model, including the Chinese ones, named a single China-domestic search vendor in any language: 0% across the board. The categories whose answers rarely cite access constraints generally shifted less.
Stripe is the most language-stable vendor in our data: 100% of English payments answers for every model, and 90-100% in Chinese.
Claude in Chinese: the same warnings, half the domestic picks
Claude flags the same broken-in-China tools the Chinese models flag, names China-domestic vendors at half GLM's rate (23.9% vs 49.3%), and never budges on LLM APIs.
Where it converges
Chinese SMS picks: Claude 54% domestic vs their 55-56%. Payments: 36% vs 33-39%. Where answers cite the same access constraints, all three models cite them at similar rates.
Where it re-orders less
On Chinese maps prompts, Kimi and GLM promote Amap to the top pick (63%, 80%). Claude keeps Google Maps first (70%) and adds Amap beside it (53%). It appends the Chinese option; they lead with it.
Where it never moves
LLM APIs. Kimi and GLM answer the Chinese question with DeepSeek (80%, 72%). Claude's Chinese LLM picks are 0% Chinese: Anthropic, all 30 of 30 answers.
Overall, Claude names a China-domestic vendor in 23.9% of Chinese answers against Kimi's 32.5% and GLM's 49.3%, and it splits advice by market half as often (28% vs ~58%). The same China constraints appear in its answers. Claude just holds its own defaults longer, and on the LLM question it does not move at all.
Chinese prompts get market-segmented answers
Share of answers that explicitly branch “domestic users” vs “overseas users” before recommending.
Sonnet 5
Kimi K3
GLM-5.2
All but one prompt are geography-free (a hosting question says “expanding to Asia”). The Chinese models respond to Chinese by splitting their advice by market in nearly 60% of answers; Claude does it in 28%. The models do more than translate. They answer as if the question were “what is the best tool for someone like you.”
We looked for a Claude-specific signature and did not find one
A natural reaction to the English convergence is to ask whether the Chinese models trained on Claude's outputs. We ran three checks: pairwise agreement is symmetric rather than Claude-tilted (Kimi matches Claude 62.7% and GLM 62.1% of the time), the disagreement pairings show no Claude inflation (p ≈ 0.35), and on 48 prompts where Claude and OpenAI's Codex reliably picked different tools, the Chinese models matched Codex about as often as Claude and carried Codex's signature pick (Statsig) at near-Codex rates. None of the three shows a Claude-specific recommendation signature.
What that does and does not establish, including the explanations these tests cannot separate and the power limits at this sample size, is in the full report's distillation section.
What this means if you build devtools
Your English mindshare travels further than you thought
Kimi and GLM recommend Vercel, Stripe, PostHog, and Clerk at Claude-like rates. A vendor that wins the English-language default wins it inside China's coding agents too, for their English-speaking users. In our runs, no Chinese cloud, auth, or analytics vendor displaced the Western leader even once in English.
The Chinese-language answer is a separate market
The same model that gives you 96% Western picks in English gives Chinese-speaking developers a different consideration set, with domestic vendors in up to half of answers. If China matters to you, your AI visibility there is not measured by English-language benchmarks.
Some categories hold in both languages; some flip
Stripe and Vercel held their top spot in both languages. Twilio lost its Chinese top spot to Aliyun SMS in two of three models, and Google Maps lost to Amap. In categories where answers cited access or payment constraints, the Chinese-language top pick often moved to a domestic vendor (payments is the exception: Alipay and WeChat Pay get added while Stripe keeps the top spot); in categories without those constraints, it usually did not.
Methodology
We ran 120 prompts (60 English, and the same 60 translated into natural developer Chinese) against 5 project repos across 12 tool categories, twice per agent, on July 25, 2026: 1,860 responses, with 0 runtime failures (42 successful runs gave no recommendation). Prompts never name a vendor; one hosting prompt mentioned expanding to Asia, and the rest are geography-free. Categories were selected for having at least one credible China-domestic alternative to the Western default.
Each model ran inside its maker's own coding agent: Claude Sonnet 5 in Claude Code, kimi-k3 in Moonshot's Kimi Code CLI (thinking effort at its default, max), and GLM-5.2 in Z.ai's ZCode CLI. This mirrors how each vendor ships its model to real users, and it means agent harness and model are confounded: differences between agents include whatever each CLI adds in system prompts and tools. Web access also differed (Claude Code and ZCode expose web tools; our Kimi profile had none). All agents ran in isolated profiles with per-prompt state wipes, in anonymized working directories, with benchmark-identifying environment variables stripped, and all API traffic originated from US IP addresses: whatever the serving side could infer from the connection pointed to the United States, and the intended experimental difference within each matched pair was the prompt language.
Primary picks and alternatives were extracted manually per our written extraction policy (a pick is what the response recommends, not the first tool named; clarifying questions without a recommendation count as non-answers, 0-5.5% per cell). Chinese responses were read in full and mapped to the same canonical vendor names as English ones. Answers that branch by market produce co-primary picks for each branch lead, which is why Chinese cells average more picks per response (2.1 vs 1.5); we report response-level rates alongside pick-share rates for that reason. Vendor classification separates China-domestic vendors from Chinese-founded global companies (TiDB, Agora, DeepSeek), which behave like Western vendors in our data.
Caveats: one suite, on five specific repos, at one point in time (mid-July 2026, days after kimi-k3's release). Per-category cells are n=10-30 per model per language, so category-level percentages carry meaningful sampling noise; the headline aggregates are n=289-734. The Chinese prompts are our translations into Simplified Chinese (the script of mainland China and Singapore; Taiwan and Hong Kong write Traditional Chinese), written to read as natural developer Chinese but not sourced from native speakers. Repo context shapes picks; these numbers describe these stacks, not global market share.
Related: What Claude Code Actually Chooses, What Codex Actually Chooses.
Run stability (same top pick, run 1 vs run 2): Sonnet 5 79.3% · Kimi K3 70.4% · GLM-5.2 65.1%.
We track what AI coding agents recommend
Amplifying benchmarks what Claude Code, Codex, Cursor, and now China's coding agents pick across thousands of prompts, and shows vendors where they win, where they lose, and what moved. New studies land here first.