Summary
- Domestic models carry the volume, and US models are kept for the hardest coding. Doubao alone averaged 180T tokens a day in June 2026 (界面新闻). IDC gives Volcano Engine 49.5% of enterprise MaaS, while Frost & Sullivan ranks Qwen first at 32.1% (每日经济新闻/IDC, 每日经济新闻/Frost). Developer threads, which are individual users' claims, still rank Claude and GPT first for complex coding and call DeepSeek the value pick (知乎, V2EX).
- Reaching Claude and GPT means using grey channels, which got pricier and riskier in 2026. Officially, OpenAI cut China off in July 2024 and Anthropic banned Chinese-controlled firms in September 2025 (Rest of World, Anthropic). Spring relays sold Opus output at about 3–28% of list price, according to forum posts. After a June ban wave, the cheapest Claude channels rose from 0.07–0.08x to 0.35x, while GPT relays sit near 4% of list (linux.do, linux.do, Tencent Cloud dev community).
- The domestic price war flipped into a wave of price rises. Zhipu's GLM Coding Plan Pro went from ¥149 to ¥538 a month in July, on top of a February rise of 30% or more, and its API prices are 83% above end-2025 (网易科技, IT之家). Baidu and Tencent raised cloud prices, and Alibaba scrapped its ¥40 Bailian plan and rations the ¥200 tier (澎湃新闻, Alibaba Cloud).
- DeepSeek is the one force pushing prices down. V4.1 Flash costs $0.30/$1.20 per 1M tokens at peak and half that off-peak, against $5/$25 for Claude Opus 5 (Artificial Analysis, kimmonismus). Even its peak V4 Pro output price of ¥27 per 1M is about a quarter of Kimi K3's ¥100 (Tencent Cloud dev community, 百度爱企查).
- Compute scarcity is the root cause. High-end GPU rents rose about 40% in six months, an H200 now rents for more than ¥66,000 a month, and 4090s listed at ¥1.32–2/hr are often unavailable, according to user posts (CSDN, 知乎). Big firms bought more than 1M gaming and workstation GPUs in 2026 and ordered $35B of RTX PRO 5500s, of which perhaps only 5–10% will be delivered (TrendForce/FT, Reuters).
- Domestic coding tools win on price, access and compliance rather than quality. Trae has about 6 million registered users, and Qoder CN (formerly Tongyi Lingma) costs ¥59 a month (极客公园, 财联社). V2EX users still rank Claude Code and Codex first, according to forum posts, and Alibaba moved staff onto Qoder only after it found the tracking code in Claude Code (V2EX, TNW).
Verdict
Chinese developers run a two-tier stack. Domestic models handle most of the volume: DeepSeek for cheap bulk calls, GLM, Kimi and Qwen for domestic coding, and Doubao at enterprise scale. Claude and GPT are still the first choice for hard coding work, and developers reach them through relays and borrowed accounts that became more expensive and more fragile through 2026. The bigger story is upstream. GPUs are scarce, and that scarcity turned 2025's price war into 2026's price rises and rationing, with DeepSeek the main exception.
Confidence is high on the structure and on official prices, which come from price pages, filings and press. It is medium on relay prices, which are forum self-reports and vendor marketing. It is low on how model filing (备案) and censorship shape daily use, because this research pass did not source those specifics.
The China AI market, from scratch
Note
A primer for readers outside China. It covers what is blocked, who sells tokens, how relay "multipliers" and coding plans work, and why GPUs set the price. One developer's monthly bill runs through all seven parts, using the same numbers as the report.
1. Two separate markets
On 9 July 2024 OpenAI stopped serving its API to developers in mainland China (Rest of World). On 5 September 2025 Anthropic stopped serving any company more than 50% owned or controlled by Chinese entities, wherever that company is based (Anthropic). Separately, many foreign tools need a VPN to work from inside China. A Tencent comparison, for example, lists Cursor Pro as needing one (Tencent Cloud dev community).
The result is two markets. US frontier models are reachable only through intermediaries. Everything else is domestic, priced in yuan and paid with local payment methods.
2. Who sells tokens in China
| Seller type | Examples | Rough Western analogue |
|---|---|---|
| Labs with their own API | DeepSeek, Zhipu (GLM), Moonshot (Kimi), MiniMax | Anthropic or OpenAI APIs |
| Hyperscaler model platforms | Alibaba Bailian (Qwen plus others), ByteDance Volcano Engine/Ark (Doubao plus others), Baidu, Tencent | AWS Bedrock, Vertex |
| Independent aggregators | SiliconFlow | Together, OpenRouter |
| Relays (中转) | Hundreds of small resellers of Claude and GPT access | No legal analogue |
The clouds resell their rivals' models. Volcano's coding plan pool includes Kimi-K3, GLM-5.3, DeepSeek-V4 and MiniMax-M3 next to Doubao (知乎).
3. The unit: yuan per million tokens
Prices are quoted in RMB per 1M tokens. This report converts at about ¥7.2 per dollar, so Claude Opus's $25 per 1M output list price is about ¥180 (linux.do).
Several providers now use time-of-day pricing, like electricity tariffs. DeepSeek's peak hours are 09:00–12:00 and 14:00–18:00 Beijing time, and off-peak calls cost half (Tencent Cloud dev community).
The worked example: Li, a developer in Hangzhou, runs a coding agent that produces 10M output tokens a month. The table counts output only, for simplicity. Agent workloads are input-heavy, which is why cache pricing and flat plans matter (part 5).
4. Relays and the "multiplier"
Relays sell site credit, usually around ¥1 per $1 of credit. Each model then burns credit at a multiplier of the official price. One October vendor charges 1.5x for Claude and 0.3x for GPT (linux.do vendor post). At 1.5x, a 1M-token Opus output costs $37.5 of credit, about ¥37.5, or about 21% of the ¥180 list price (an estimate, converting at ¥7.2/USD).
Relays can sell this far below list because they resell flat-rate subscriptions. A $200-a-month Claude Max seat produced about $2,700 of API-equivalent usage in January 2026, and $3,500–5,200 after Opus got a 1M-token context window, according to the linux.do author's estimate (linux.do). It works like buying wholesale from someone with an all-you-can-eat membership.
| Li's route (10M output tokens/month) | ¥ per 1M output | Monthly bill | Share of Claude list |
|---|---|---|---|
| Claude Opus, official ($25) | ~¥180 | ~¥1,800 | 100% |
| GPT-5.6 Sol, official ($20) | ~¥144 | ~¥1,440 | 80% |
| Kimi K3, first-party API | ¥100 | ¥1,000 | 56% |
| Claude via relay, Oct 2026 (1.5x) | ~¥37.5 | ~¥375 | ~21% |
| Doubao-Seed-Evolving | ¥30 | ¥300 | 17% |
| GLM-5.3 on SiliconFlow | ¥28 | ¥280 | 16% |
| DeepSeek V4 Pro, peak | ¥27 | ¥270 | 15% |
| Claude via relay, spring 2026 (typical) | ~¥20 | ~¥200 | ~11% |
| DeepSeek V4 Flash, peak | ¥9 | ¥90 | 5% |
| GPT via relay, Oct 2026 (0.3x) | ~¥6 | ~¥60 | 3% (4% of GPT list) |
| DeepSeek V4 Flash, off-peak | ¥4.5 | ¥45 | 2.5% |
5. Metered pricing versus coding plans
A coding plan is a flat monthly fee capped by requests or credits, the domestic version of Claude Max. Alibaba's Bailian Pro costs ¥200 a month for up to 90,000 requests, bundling Qwen, GLM-5, Kimi and MiniMax-M2.5 for use inside Claude Code (Alibaba Cloud).
For Li, ¥200 flat beats ¥280 metered GLM-5.3 if he stays under the caps. GLM's own Pro plan, at ¥538 since July, does not. Kimi's coding access starts at ¥99 (知乎). Flat plans are where the 2026 price rises bit hardest (section 3).
6. Why GPUs set the price
Chinese providers cannot freely buy top Nvidia chips. The H20, Nvidia's China-specific datacenter chip, is effectively off the table (section 4). Providers are left with domestic accelerators, workstation cards and consumer cards like the RTX 4090.
When demand for tokens outruns GPUs, providers raise prices or ration capacity. Zhipu's cost of sales rose 213% on compute fees (网易科技). The same scarcity reaches Li directly if he wants to rent a 4090 for fine-tuning.
7. Model filing (备案)
Generative-AI services offered to the Chinese public go through a regulatory filing before launch. That is one reason domestic vendors market themselves as "compliant". This pass did not source the specifics or measure the effect on what developers can build, so the report leaves it as an open question.
In one line: the blocks split the market in two, relays put the US frontier at roughly 4–21% of list price for those who accept ban risk, and GPU scarcity is why Li's domestic bill went up in 2026 everywhere except at DeepSeek.
| Value | Change | |
|---|---|---|
| Doubao daily tokens (Jun 2026) | 180T | ▲ 10x+ YoY |
| GLM Coding Plan Pro | ¥538/mo | ▲ +261% in July |
| DeepSeek V4.1 Flash peak output | $1.20/1M | unchanged |
| Cheapest Claude relay multiplier | 0.35x | ▲ from 0.07–0.08x |
| H200 rent per card | ¥66k/mo | ▲ high-end rents +40% in 6 months |
| Chinese share of OpenRouter tokens | 54%+ | ▲ from <15% a year earlier |
1. What developers actually use, and for what
The community view is consistent across buyer guides and forums. Developers use Claude (Opus or Sonnet 4.6–4.7) and GPT-5.x for complex coding and refactors. DeepSeek V4 is the "value king" for high-volume calls. GLM-5.x, Kimi K2.6/K3 and Qwen3.6+ are the domestic coding alternatives, with better Chinese-language fit. One Zhihu tester called DeepSeek-V4 "far ahead" on cost-performance for workloads of a million tokens a day, which is one user's claim (知乎). TNW also notes that Claude stays popular with Chinese developers who use workarounds (TNW).
| Model family | Typical job, per developers | Evidence | Source |
|---|---|---|---|
| Claude Opus/Sonnet | Hardest coding, refactors | Anecdotal | 知乎, TNW |
| GPT-5.x / Codex | Hard coding. One poster ranks GPT above DeepSeek and avoids Claude because "it banned all my accounts" | Anecdotal | V2EX |
| DeepSeek V4 / V4.1 | Bulk and agent calls ("fast, cheap, smart enough") | Anecdotal | V2EX, 知乎 |
| GLM-5.x, Kimi K3, Qwen | Domestic coding, often inside coding plans | Data (plan bundles) plus anecdotal | Alibaba Cloud, 知乎 |
| Doubao (Seed) | Enterprise volume via Volcano Engine | Data (company claim) | 界面新闻 |
| MiniMax M2.5/M3 | Present in Bailian and Volcano coding pools | Data. No usage share found | Alibaba Cloud, 知乎 |
The volume numbers are data, but they disagree on who leads.
| Measure | Value | Source |
|---|---|---|
| Doubao average daily tokens | 180T (June 2026), up from 120T+ in March, more than 10x year on year | 界面新闻 |
| China public-cloud LLM calls, 2025 (IDC) | 1,944T tokens, 16x the 114T of 2024. Volcano 49.5% enterprise MaaS share, up from 46.4% | 每日经济新闻 |
| Enterprise LLM usage, H2 2025 (Frost & Sullivan) | 37.0T tokens/day, +263% vs H1. Qwen first at 32.1%, up from 17.7% | 每日经济新闻 |
| DeepSeek, one DeepInfer cluster | 155T tokens/day, $38 revenue per GPU-hour off-peak (company claim). A $1B ARR claim is unverified | Yuchen Jin, X |
| Six leading Chinese AI firms' revenue | About 10% of OpenAI plus Anthropic AI revenue, Epoch AI estimates | Epoch AI |
IDC and Frost measure different scopes, so "who leads China" is contested. Volcano leads on cloud MaaS calls, and Qwen leads on Frost's enterprise-usage measure.
Global context only. OpenRouter data shows Chinese models' share of global tokens climbing from under 15% to over 54% in a year, led by DeepSeek (CNN). Chinese models have held at least 30% of OpenRouter enterprise tokens every week since 8 February 2026, peaking at 46% (CNBC). These figures mostly reflect developers outside China, so they say little about what Li uses.
2. Reaching Claude and OpenAI: the grey market in brief
The separate relay long read covers this market in depth. This section is market-level only.
The blocks, as a timeline (all data except where noted):
| Date | Event | Source |
|---|---|---|
| 9 Jul 2024 | OpenAI cuts China API access. Azure's China JV (21Vianet) reported unaffected | Rest of World |
| 5 Sep 2025 | Anthropic bans firms more than 50% Chinese-owned or controlled. Reports in July 2026 say this was extended to majority-owned subsidiaries | Anthropic |
| Feb 2026 | Anthropic names DeepSeek, Moonshot and MiniMax: about 24,000 fraudulent accounts, more than 16M exchanges for distillation | Anthropic |
| Mar–Apr 2026 | Hidden detection code ships in Claude Code from v2.1.91 (2 April). It checks for Shanghai/Urumqi timezones and Chinese proxy domains | TNW |
| 10 Jun 2026 | Anthropic tells the Senate that Qwen-affiliated operators ran about 25,000 accounts and 28.8M exchanges from April to June. Alibaba denies it | TNW |
| 12–30 Jun 2026 | US Commerce orders Claude Fable 5 and Mythos 5 turned off for all foreign users. Access returns 1–2 July | 9to5Mac |
| Late Jun 2026 | Claude Code ban wave hits mainland users, including those on proxies and relays (counts anecdotal) | Tencent Cloud dev community |
| 1 Jul 2026 | An Anthropic engineer says the detection "experiment" was removed | TNW |
| Early Jul 2026 | Alibaba bars employees from Claude Code and points them to Qoder. The notes date this 10 July, but TNW's report is dated 3 July, so the exact day is uncertain | TNW |
The detection. Two independent Chinese write-ups reverse-engineered the code (Tencent Cloud dev community, Aliyun dev community). It combined the system timezone with an obfuscated blocklist of 147 domains matched against the custom API endpoint setting. Flagged users were marked steganographically in the system prompt, through date formats and apostrophes. The list included .cn domains, Baidu, Alibaba and ByteDance domains, DeepSeek and Zhipu, and "mostly various API relay services".
The presence of DeepSeek and Zhipu domains implies something worth noting (this is an inference). A meaningful number of Chinese developers run the Claude Code harness against domestic model endpoints, not only against Claude. Bailian's plan documentation openly targets that use (Alibaba Cloud).
The channels (market level, from the linux.do relay analysis; anecdotal unless noted):
- Plan arbitrage. Relays resell $200 Claude Max seats, where cache reads are free.
- Cheaper, dirtier channels. These are said to use stolen cards, scraped organization seats, or reverse-engineered Cursor, Antigravity, GitHub Copilot or Kiro access (linux.do).
- VPN plus overseas subscriptions. Cursor Pro at $20 a month needs a VPN, for example.
- Domestic "no relay" offers. Tencent's WorkBuddy international edition at $10 a month, and AWS Bedrock for enterprises. This comes from a Tencent promotional article, so it is anecdotal (Tencent Cloud dev community).
Prices against official list:
| When | Channel | Price | Share of official | Evidence |
|---|---|---|---|---|
| Mar–Apr 2026 | Claude Code relays, Opus 4.6 output | ¥5–50 per 1M, mostly ¥10–35 | ~3–28% (derived) | Anecdotal (linux.do) |
| Mar 2026 | Relay cost of $1 of official-price Claude | ¥0.2–0.4, or ¥0.6–0.7 all-in, the linux.do author estimates | n/a | One user's claim (linux.do) |
| Oct 2026 | Cheapest "reverse-Kiro" Claude | 0.35x, up from 0.07–0.08x. Genuine channels 0.8–1x | Site multipliers, not comparable across sites | Anecdotal (linux.do) |
| Oct 2026 | One vendor: Claude / GPT | 1.5x / 0.3x of official, at ¥1 per $1 of credit | ~21% / ~4% | Vendor marketing (linux.do) |
Reliability. In spring, some vendors paused sign-ups or raised prices 1.5–3x (linux.do). In October, posters blame the Claude price rise on ban costs. The original poster in one thread lost a dozen-plus accounts, and Anthropic does not refund banned accounts. Others say most relays are not "full-strength" Claude Max, and that GPT relays are often "watered down" (降智), according to forum posts (linux.do).
3. The price war that turned into a price-rise wave
The Paper called 2026 a "price-rise wave" for Chinese AI (澎湃新闻). The timeline below is all data.
| Date | Move | Source |
|---|---|---|
| 12 Feb | Zhipu GLM Coding Plan up at least 30%, first-purchase discounts removed | IT之家 |
| 11 Mar, 9 May | Tencent Cloud price rises | 澎湃新闻 |
| 16 Mar / 8 Apr | GLM-5-Turbo +20% / GLM-5.1 +10%. GLM-5.1's cached coding-token price is now near Claude Sonnet 4.6 | 澎湃新闻 |
| 20 Mar / 13 Apr | Bailian ¥40 Lite closed to new buyers / to renewals | Alibaba Cloud |
| 18 Apr | Baidu AI Cloud compute prices +5–30%. Alibaba Cloud also raises AI compute prices | 澎湃新闻, CSDN |
| Until 5 May | DeepSeek V4-Pro launch at 75% off: ¥0.025 cached input, ¥3 input, ¥6 output | 澎湃新闻 |
| Late Jul | GLM Coding Plan repriced: ¥118/¥538/¥1,078, now credit-based with peak weighting | 网易科技 |
| 17 Aug | DeepSeek moves to peak/off-peak pricing. Peak output up to ¥9 (V4 Flash) and ¥27 (V4 Pro) | Tencent Cloud dev community |
| Sep | Kimi rebuilds membership as ¥49/¥99/¥199/¥699 tiers | 知乎 |
| Before | After | |
|---|---|---|
| Lite | 49 ¥ | 118 ¥ |
| Pro | 149 ¥ | 538 ¥ |
| Max | 469 ¥ | 1,078 ¥ |
Why Zhipu raised prices, from its filings. Its 2025 results show more than 242,000 paying Coding Plan developers, API prices 83% above end-2025, and demand still above supply. Cost of sales rose 213.3% to ¥427.7M against revenue of ¥724.3M and R&D spending of ¥3.18B. Before the July change the plan was chronically sold out. Existing subscribers kept their old prices, and domestic prices now roughly match Z.ai's overseas tiers of $18/$80/$168 (网易科技).
What metered tokens cost now (per 1M tokens, RMB):
| Model / venue | Input | Output | Notes | Source |
|---|---|---|---|---|
| DeepSeek V4.1 Flash, first-party | $0.30 | $1.20 | Peak. Off-peak half | Artificial Analysis |
| DeepSeek V4 Pro, first-party | – | ¥27 | Peak. Off-peak half | Tencent Cloud dev community |
| DeepSeek-V4.1-Flash, SiliconFlow | ¥2 | ¥8 | Cache ¥0.04. Half price 02:00–08:00 | SiliconFlow |
| GLM-5.3 / GLM-5.3-Flash, SiliconFlow | ¥8 / ¥0.8 | ¥28 / ¥2.8 | GLM-5.3 cache ¥2 | SiliconFlow |
| Kimi-K2.7-Code, SiliconFlow | ¥6.5 | ¥27 | SiliconFlow | |
| Qwen3.8-27B / LongCat-2.0, SiliconFlow | ¥3 / ¥5 | ¥12 / ¥20 | SiliconFlow | |
| Kimi K3, first-party | ¥20 (¥2 cached) | ¥100 | 2.8T open weights, released 17 July. Secondary aggregator | 百度爱企查 |
| Doubao-Seed-Evolving, Volcano | ¥6 | ¥30 | 1,024k context. Cheaper tiers from ¥0.15 input | 火山引擎 |
| RMB per 1M output | |
|---|---|
| GPT relay 0.3x | 6 ¥ |
| DeepSeek V4 Flash peak | 9 ¥ |
| Kimi-K2.7-Code (SiliconFlow) | 27 ¥ |
| DeepSeek V4 Pro peak | 27 ¥ |
| GLM-5.3 (SiliconFlow) | 28 ¥ |
| Doubao-Seed-Evolving | 30 ¥ |
| Claude relay 1.5x | 37.5 ¥ |
| Kimi K3 first-party | 100 ¥ |
| GPT-5.6 Sol official | 144 ¥ |
| Claude Opus 5 official | 180 ¥ |
At peak, DeepSeek's output price is about 17–21x cheaper than the US flagships, and Kimi K3's output is about 11x DeepSeek V4 Flash's peak price (both ratios derived) (kimmonismus, 百度爱企查).
Coding plans (monthly, data unless noted):
| Plan | Price | Limits and quirks | Source |
|---|---|---|---|
| GLM Coding Plan | ¥118 / ¥538 / ¥1,078 | Credits weight input, cache, output, tools and peak use | 网易科技 |
| Bailian Coding Plan Pro | ¥200 (first month ¥39.90) | 6,000 requests per 5h, 45,000 a week, 90,000 a month. Daily allotments released at 09:30 | Alibaba Cloud |
| Volcano Ark Coding Plan | ¥40 / ¥200 (promos ¥9.9 / ¥49.9) | Doubao, Kimi-K3, GLM-5.3, DeepSeek-V4, MiniMax-M3 | 知乎 |
| Kimi membership | ¥49 / ¥99 / ¥199 / ¥699 | Code from ¥99 (¥79/month paid annually) | 知乎 |
User complaints from February, from one blog post (博客园):
- Volcano's ¥40 plan was described as oversold and slow, with frequent 400/429 errors.
- After Zhipu's rise, GLM-5 burned quota at 3x during peak hours and 2x off-peak.
- Kimi's plan had limited tool compatibility.
- Entry tiers on both Volcano and Bailian were about 1,200 requests per 5 hours and 18,000 a month.
Free tiers have shrunk.
- SiliconFlow's free models are now small ones such as Hunyuan-MT-7B and PaddleOCR-VL (SiliconFlow).
- Volcano gives new users 500,000 free tokens per language model (火山引擎).
- The free options that remain sit mostly in the IDEs (section 5).
Local deployment is a hobbyist hedge, according to user posts.
- A modded 48GB RTX 4090 running a Qwen3.8 Flash model reportedly decodes at about 120 tok/s single-stream.
- Unsloth's 1-bit GGUF of DeepSeek V4.1-Flash-1017 shrinks it from 508GB to 250GB, enough for machines with 128GB or more of RAM.
- DeepSeek open-sourced its DeepInfer serving stack on 20 October.
- Critics note that quantized local models at about 30 tok/s trail cloud Claude in quality (MinLiBuilds, X).
- A used 4090 sold for about ¥15,000 in June (搜狐).
4. Compute for builders: renting GPUs
For a fine-tuning or serving builder, the 4090 is the default card. Its price is low on paper, but it is often hard to actually get one.
| Low | High | |
|---|---|---|
| C2C sharing marketplaces | 0.8 ¥/hr | 1 ¥/hr |
| Specialist rental (AutoDL, 智星云) | 1.32 ¥/hr | 2 ¥/hr |
| Big clouds (Alibaba, Tencent, Huawei) | 3 ¥/hr | 8 ¥/hr |
| Item | Value | Evidence | Source |
|---|---|---|---|
| 4090 on specialist sites | ¥1.32–2.0/hr. 智星云 rose from ¥1.32 to ¥1.50 in May | Self-reported list prices | 知乎 |
| 4090 availability | Often unavailable "even with money in hand". Spot pools squeezed by SME long-term leases and big A-share compute contracts. Modded 48GB cards rentable on some platforms (晨涧云) | Anecdotal | 搜狐 |
| High-end rents | Up about 40% in six months (Hurun) | Data (secondary) | CSDN |
| H200 / H100 monthly rent | More than ¥66,000 per card (about $12.7/hr, est.) / ¥55–60k, up 20% | Data (secondary) | CSDN |
Why the H20 route is closed:
- In September 2025 Beijing told major firms to stop buying H20 and RTX Pro 6000D, through the internet regulator (the CAC).
- In October 2025 customs cracked down on Nvidia imports at ports.
- From 15 January 2026 the US Commerce Department's export office (BIS) moved H200 and MI325X licences for China to case-by-case review, with a 25% tariff and a 50% volume cap.
- Licensed H200 volumes are reportedly small and mostly kept in Hong Kong (The Frontier).
The workaround is workstation and gaming silicon. Chinese firms bought more than 1 million Nvidia gaming and workstation GPUs in 2026, mostly RTX Pro 6000D, RTX Pro 5000 and RTX 5090. Domestic chip output more than doubled but still falls short. Combined Alibaba, ByteDance and Tencent capex is expected to roughly double to about ¥700bn ($104bn) (TrendForce/FT). The RTX Pro 6000D shows up on both the 2025 "stop buying" list and the 2026 purchase list, so the notes leave it unclear how strictly the 2025 instruction holds.
ByteDance, Alibaba and Tencent have also ordered more than 3 million RTX PRO 5500 cards. These are 84GB, US-compliant cards at about $11,400 each, mainly for inference, worth over $35B in total. Nvidia cannot guarantee delivery, and some buyers may get only 5–10% of what they ordered (Reuters). DeepSeek's DeepInfer supports both Blackwell and Huawei Ascend, which hedges the same problem in software (Yuchen Jin, X).
The link to token prices is explicit. Baidu and Tencent cite compute demand and hardware costs for their price rises (澎湃新闻), and Zhipu's cost of sales tripled on compute fees (网易科技).
5. Tools: Chinese coding agents and IDEs
| Tool | Vendor | Price | Scale | Community verdict (user claims) | Sources |
|---|---|---|---|---|---|
| Trae / Trae CN | ByteDance | Free (CN) | About 6M registered users, 1.5M+ MAU, about 100B lines generated (Dec 2025). More than 80% of ByteDance engineers use it | Free and feature-complete, but slow "thinking" | 极客公园, V2EX |
| Qoder CN (Tongyi Lingma, renamed 20 May 2026) | Alibaba | Free 100 credits a month. Pro ¥59. Enterprise ¥99 per seat-month | Alibaba's internal default after the Claude Code ban | Thorough but slow. One user's ¥200 quota ran out fast, "several times more expensive than Cursor" | 财联社, V2EX |
| Qoder (international) | Alibaba | Pro $20 (about ¥142). Teams ¥300 per seat-month with 3,000 credits (from 30 April) | – | – | 财联社 |
| CodeBuddy CN / WorkBuddy | Tencent | WorkBuddy international $10 a month (promotional) | – | Updates faster than Trae CN | V2EX, Tencent Cloud dev community |
| Comate | Baidu | Not sourced | Not sourced | Not covered in this pass | – |
| Claude Code, Codex, OpenCode | Anthropic, OpenAI, open source | Via relays or overseas accounts | – | Most common "best overall" answer | V2EX |
The V2EX thread from May 2026 has about 50 replies. It is a sample, not a survey, but it fits the buyer guides (V2EX, 知乎). Domestic vendors pitch "compliance, low domestic latency, no VPN, fit with domestic stacks (UniApp, mini-programs, Spring Boot)" rather than raw quality (TNW, citing CSDN).
The most interesting pattern for harness builders is the split between harness and model. Domestic coding plans are sold for use inside Claude Code (Alibaba Cloud), and Anthropic's blocklist targeted Claude Code pointed at Zhipu and DeepSeek endpoints (Tencent Cloud dev community). Many developers keep the US harness and swap the model underneath. This is an inference from the two sources, not a measured share.
6. Culture: complaints, compliance and sentiment
What developers complain about (user claims):
- Ban fatigue. "It banned all my accounts" (V2EX). A dozen-plus lost accounts with no refunds (linux.do).
- "Dumbed-down" (降智) relays that don't serve the model they claim to (linux.do).
- Oversold plans. 429 errors, quota multipliers at peak, and daily 09:30 allotments (博客园, Alibaba Cloud).
- The GLM rise as a "horror story". Users asked why they should pay ¥538 a month for GLM Pro instead of subscribing to GPT or Claude directly. Defenders cited RMB payment, domestic network access, easier procurement and data compliance for institutions that cannot reliably buy foreign models (网易科技).
Employer policies.
- Alibaba's ban is the landmark case (TNW).
- Security vendors frame it as a data-compliance issue. Security firm Huorong said the hidden tracking raised cross-border data-compliance concerns, not just transparency ones.
- Outside analysts expected it. Asia Society's Lizzi Lee said it is unsurprising that big Chinese firms bar a US tool that can detect Chinese use (TNW).
- ByteDance standardizes on its own tool. More than 80% of its engineers use Trae internally (极客公园).
Censorship and filing (备案): unresolved. This pass found no sourced specifics on how model filing or content rules change developers' daily model choice. The sourced compliance story is about data sovereignty and procurement, not content filtering.
Policy is tightening on both sides (reported, not final).
- China. On 16 October China proposed export controls on 23 AI technologies, including large models, agents and compute clusters. Overseas deployment of restricted models would need approval. APIs, apps, SaaS and already-public open source are exempt.
- United States. DHS is reportedly requiring civilian agencies to remove DeepSeek products (choblin29, X). The Fable 5 suspension also showed that US policy can switch off frontier access for all foreigners, not only Chinese users (9to5Mac).
Sentiment, in short:
| US models (Claude, GPT) | Domestic models | |
|---|---|---|
| Pull | Best for hard coding. Claude Code and Codex are seen as the gold standard | Cheap (especially DeepSeek), RMB billing, no VPN, compliant |
| Push | Bans, relay degradation, rising Claude relay prices, tracking distrust | Price rises, rationing, slower tools, quota burn |
| Trend | GPT gaining as the cheaper grey-market option | Defaulting at employers. DeepSeek is the cost anchor |
What to watch
- DeepSeek's peak price. Watch whether V4.1 Flash holds at $1.20 per 1M output ($0.30 input) now that DeepInfer is open source, and whether rivals' daytime rates (GLM-5.3 at ¥28) follow it down (Artificial Analysis, SiliconFlow).
- Claude relay multipliers. The cheapest is 0.35x, up from 0.07–0.08x. A fall back below 0.1x would mean ban costs eased, and a rise toward 1x would mean the grey market is pricing in permanent bans (linux.do).
- Coding-plan rationing. Watch whether Bailian Pro's 09:30 daily allotments keep selling out at ¥200, whether Kimi or Volcano raise prices toward GLM's ¥538 Pro tier, and Zhipu's next paying-developer count (last 242,000) (Alibaba Cloud, 网易科技).
- RTX PRO 5500 deliveries. Watch actual fill against the 3M+ unit, $35B order, where 5–10% is the bear case (Reuters).
- Builder rents. A 4090 above ¥2/hr on specialist sites, or an H200 above ¥66,000 a month, would mean scarcity is still worsening (知乎, CSDN).
- Final rules. Watch the final text of China's 23-technology export-control proposal and any further Anthropic ownership restrictions after the July subsidiary extension (choblin29, X, Anthropic).
Sources
- Artificial Analysis (X), 2026-10-19: DeepSeek V4.1 Flash 1017
- kimmonismus (X), 2026-10-17: DeepSeek V4.1 Flash-1017 pricing vs GPT/Claude
- Yuchen Jin (X), 2026-10-20: Quoting DeepSeek DeepInfer blog
- Epoch AI, 2026-10-06: How do Chinese AI companies make money?
- The Next Web, 2026-07-03: Alibaba bans Claude Code after Anthropic is caught tracking Chinese users
- 9to5Mac, 2026-07-01: Claude Fable 5 cleared to return as US lifts export control restriction
- Tencent Cloud Developer Community, 2026-07-01: 终于，Claude Code 封号的原因被曝光了
- Aliyun Developer Community, 2026-07-01: Claude Code 封号真实原因曝光
- Anthropic, 2026-02-23: Detecting and preventing distillation attacks
- Rest of World, 2024-06-25: OpenAI to cut off API access in China on July 9
- Tencent Cloud Developer Community, 2026-07-12: 不用中转、不怕封号：3 个稳定用上 Claude 的方案对比
- LINUX DO, 2026-04-09: 用过的Claude Code中转商报价分享和成本分析
- LINUX DO, 2026-10-04: 现在 claude 有靠谱实惠的中转吗？
- LINUX DO, 2026-10-12: 客优 Claude / GPT 倍率说明
- CNN, 2026-09-26: The US wants the world to pick a side on AI
- CNBC, 2026-09-26: Chinese AI models surge in global popularity — and Washington is worried
- 界面新闻, 2026-06-23: 豆包大模型日均Tokens调用量达180万亿
- 每日经济新闻, 2026-05-07: IDC：2025年中国MaaS市场Token调用量激增16倍，火山引擎份额49.5%
- 每日经济新闻, 2026-02-24: 机构报告：中国企业调用大模型日均达37万亿tokens，阿里云千问占比第一
- 网易科技, 2026-07-31: 智谱GLM Coding Plan不用抢了，最多涨价两倍，价格向海外看齐
- IT之家, 2026-02-12: 智谱对 GLM Coding Plan 套餐价格进行结构性调整，整体涨幅 30% 起
- Alibaba Cloud Help Center, 2026-10-10: Coding Plan概述 (百炼)
- 知乎, 2026-10-07: 程序员的Coding Plan 完整指南（2026年10月7日持续更新）
- 博客园, 2026-02-25: 2026年国内主流AI Coding Plan套餐全对比｜开发者避坑指南
- 澎湃新闻, 2026-04-26: 涨价浪潮下，DeepSeek推动AI“价格战”：百万Token输入0.025元
- Tencent Cloud Developer Community, 2026-08-12: DeepSeek API涨价：8月17日起正式调整价格
- 火山引擎, 2026-10-15: AI Hub 模型广场
- 百度爱企查, 2026-10-02: kimi-k3定价
- SiliconFlow, 2026-10-20: 模型价格总览
- 搜狐, 2026-06-10: 2026年6月GPU算力平台租用4090抢卡攻略
- 知乎, 2026-07-15: 2026年GPU算力平台深度横评：6大平台真实价格与性能数据全公开
- CSDN, 2026-05-12: 2026 怎么租到高性价比GPU算力
- TrendForce, 2026-10-15: Chinese firms reportedly buy over 1M Nvidia gaming/workstation chips
- Reuters (via WSAU), 2026-10-20: ByteDance, Alibaba, Tencent place over $35 billion worth of Nvidia chip orders
- The Frontier, 2026-10-03: H200 and MI325X China licenses go case-by-case
- V2EX, 2026-05-07: AI 编程工具大家实际使用中哪个综合下来最好用？
- 财联社, 2026-04-24: Qoder Teams版新购价格调整为300RMB/席位月
- 极客公园/腾讯新闻, 2025-12-29: 一年 200 次更新，字节 TRAE 如何拿下全球 600 万开发者
- choblin29 (X), 2026-10-16: China AI export-control proposal relay
- 知乎, 2026-05-13: 2026 年 AI 编程实测：6 款顶流大模型对比
- MinLiBuilds (X), 2026-10-12: Local deployment on modded 4090 / Unsloth GGUF / DeepInfer