China AI Bench
8 min read

China vs US AI models, mid-2026: a scorecard

CommentaryBenchmarks Deskby Benchmarks Desk, Testing & rankings desk

At mid-2026 the China-US model race has a scorecard with two columns: China wins open weights, volume, and price; the US still holds the frontier lead. 13 straight weeks of Chinese token volume, an 80% OpenAI price cut, and the numbers in between.

For thirteen straight weeks, Chinese models have moved more tokens on OpenRouter than American ones. The last measured week, July 20-26, was 38.6 trillion tokens, 66.5% of the router's volume [1]. Meanwhile the American frontier still holds the highest scores on the aggregate index, and the US government's policy toward Chinese open weights flipped from accusation to exemption inside a week [2][3].

Both things are true, and the mid-2026 picture only makes sense if you hold them together. This is a scorecard of what we can verify, with the numbers dated and labeled.

The volume story

The OpenRouter weekly data is the cleanest signal we have. Chinese model token volume has now exceeded US volume for 13 consecutive weeks since late April, and the 66.5% share for July 20-26 is the latest snapshot [1]. The drivers aren't mysterious: DeepSeek's free-tier and cheap-tier models power a lot of agent traffic. OpenCode alone pushed 8 trillion tokens through DeepSeek V4 Flash on a single day, August 1 (5T on the free tier, 3T on its paid Go plan) [4].

DeepSeek's broader adoption numbers, from Presenc AI research, put weekly active users at 188M in Q1 2026, with 11,400 organizations self-hosting the weights, and DeepSeek-V4 at about 23% of open-source production AI applications [5]. Those are third-party research numbers, captured March 31, and the self-hosting figure matters most: it's the largest sovereign-AI deployment footprint of any model family.

The price war, from both sides

The pricing table as of our August 6 snapshot:

ModelInput $/1MOutput $/1MNote
DeepSeek V4 Pro0.4350.87cache hit $0.003625, 75% permanent cut [1]
DeepSeek V4 Flash0.140.28cache hit $0.0028 [6]
GLM-5.21.404.40snapshot July 31 [1]
Kimi K33.0015.00cache hit $0.30 [7]
GPT-5.6 Luna0.201.20after the July 31 -80% cut [1][8]
Claude Opus 55.0025.00half the price of the frontier flagship [9]

Read the table upside down and you get the year's actual story. OpenAI cut Luna's price 80% on July 31, three weeks after the model launched, and Sam Altman's stated goal was best price-to-intelligence at every tier [8]. That's the US responding to a pricing war China started, with DeepSeek V4 Pro's cache mechanism and cheap Chinese APIs pulling Western developers onto Chinese models [10]. Artificial Analysis estimated Luna's per-task cost on Agents Last Exam at about 1% of Fable 5's [1]. On the other side, Kimi K3 is deliberately priced at $3/$15, the most expensive Chinese model ever, and Moonshot's enterprise lead said open source doesn't have to be cheap [11]. The race has two price strategies now, and they're both working.

The capability gap

On the aggregate intelligence index, the frontier is still American. Artificial Analysis's Intelligence Index v4.1 puts Claude Fable 5 at 59.9, GPT-5.6 Sol at 58.9, and the best Chinese model, Kimi K3, at 57.1, third overall [12]. That 2-3 point gap is real and stable. NIST CAISI's independent eval of DeepSeek V4 Pro estimated the model sits about eight months behind the US frontier, a gap that narrowed from 12-14 months for V3 [13]. Our own DeepSeek V4 review covers that eval in detail.

Where China wins is specificity, not aggregate: K3 took the Frontend Code Arena top spot at 1679 [14], GLM-5.2 holds the open-source AA lead [15], and the model lineups ship faster than anything the US produces open. Hugging Face's CEO said it in a CNBC interview in early August: China is "crushing" the West on open models, and he expects Chinese tools to catch the US frontier by end of 2026 or 2027 [2].

Open weights vs API-only

The strategic divide is cleaner than the capability gap. China ships weights: DeepSeek (MIT), K3 (Modified MIT, 2.8T), GLM-5.2 (MIT), Qwen 3.8 (weights promised, tracking), MiniMax H3 (promised, tracking). The US frontier mostly ships APIs. Anthropic said it has never advocated banning open-weight models [16], and 25 companies including NVIDIA signed an open letter in late July arguing the world needs frontier open weights, not just frontier closed ones [17]. The policy backdrop is incoherent: a White House official accused K3 of industrial distillation with no published evidence [3], then Bloomberg reported Chinese open weights would be spared US safety tests [18]. We compile what's verifiable and label the rest.

Our take

The mid-2026 scorecard reads like this: China wins distribution, volume, price, and open infrastructure. The US wins the frontier aggregate and, for now, the enterprise trust layer. OpenAI cutting prices 80% is the most honest signal of the year: the US validates the China playbook by adopting it.

The thing nobody should miss is the distribution layer. OpenRouter, Ollama, OpenCode, vLLM: none of them are Chinese or American, and all of them now treat Chinese and US models as interchangeable plug-ins. When the router doesn't care where the model is from, "Chinese model" and "American model" become procurement categories, not product categories. That's the structural change, and it's why token volume flipped.

What I can't tell you is whether the aggregate gap closes on schedule. Hugging Face's CEO says end of 2026 or 2027 [2]; DeepSeek's own founder, in leaked notes from a closed investor meeting we can't fully verify, put the compute gap at 20x and the model gap at 12-18 months, targeting 3-6 [19]. Both are estimates with clear interests attached. The weekly token numbers are not estimates.

Risks and what we're watching

  • Policy whiplash. Distillation accusation one week, safety-test exemption the next; the signal to Western enterprises is "figure it out yourself," which some will read as risk and others as opportunity.
  • Price sustainability. Luna at $0.20 input is a loss-leader posture; if OpenAI reverts, the Chinese volume story gets a headwind.
  • The knowledge gap. K3 and GLM trail on HLE/GPQA by about 5% [20]; knowledge-heavy buyers are the segment where US models still win.
  • Open-weight promises. Qwen 3.8's weights and MiniMax H3's weights are both outstanding; the scorecard updates the day either lands.

Bottom line

China has won the volume race, the price race, and the open-infrastructure race so far in 2026. The US still holds the frontier aggregate and the knowledge lead. The 80% price cut from OpenAI and the 13 straight weeks of Chinese token volume are the two numbers I'd quote in any planning meeting, because they're the ones with dates and sources. Everything else is a forecast.

Sources

  1. 36Kr / First New Voice (OpenRouter 13 weeks, 66.5% share, pricing table, Luna task cost), 36kr.com/p/3919319636290946 (published 2026-07-31, captured 2026-08-06, media report of OpenRouter + official pricing)
  2. OSChina / CNBC (Hugging Face CEO: China crushing the West on open models; catch-up end-2026/2027), oschina.net/news/487177/hugging-face-china-ai-race-open-models (published 2026-08-04, captured 2026-08-06, media report of CNBC interview)
  3. Zhihu (White House distillation accusation, no evidence published), zhihu.com/question/2063537141473449457 (published 2026-07-22, captured 2026-08-06, community)
  4. OSChina / OpenCode (8T tokens single day, Aug 1), oschina.net/news/486803 (published 2026-08-04, captured 2026-08-06, media report)
  5. Presenc AI Research (188M weekly active, 11,400 self-hosted, V4 23% of open-source production apps), presenc.ai/research/deepseek-usage-statistics (published 2026-03-31, captured 2026-08-06, third-party research)
  6. Puter dev blog (V4 Flash $0.14/$0.28, cache $0.0028), developer.puter.com/tutorials/deepseek-api-pricing/ (published 2026-06-17, captured 2026-08-06, official)
  7. Ollama (K3 $3/$15, cache $0.30), ollama.com/library/kimi-k3 (published 2026-08-01, captured 2026-08-06, official)
  8. Tencent News / Wall Street CN (Altman pricing matrix, Luna -80%, Terra -20%, Sol Fast), news.qq.com/rain/a/20260731A028E600 (published 2026-07-31, captured 2026-08-06, media report)
  9. Anthropic official blog (Claude Opus 5, $5/$25), anthropic.com/news/claude-opus-5 (published 2026-07-24, captured 2026-08-06, official)
  10. AIBase (Luna price-performance overtakes V4 Pro per AA; cheap Chinese APIs as driver), news.aibase.com/zh/news/30025 (published 2026-07-31, captured 2026-08-06, third-party eval, via media)
  11. Tencent News / Jiupai (K3 pricing strategy quote), news.qq.com/rain/a/20260724A0A6XX00 (published 2026-07-24, captured 2026-08-06, media report)
  12. MACGPU Blog (AA Intelligence Index v4.1: Fable 5 59.9, Sol 58.9, K3 57.1), macgpu.com/zh/blog/2026-0722-kimi-k3-kaiyuan-quanzhong-fabu.html (published 2026-07-22, captured 2026-08-06, third-party eval, via community blog)
  13. NIST CAISI (DeepSeek V4 Pro eval, 8-month gap), covered in our DeepSeek V4 review, nist.gov (2026-05-01) [third-party eval]
  14. Jiemian (Frontend Code Arena 1679), jiemian.com/article/14786633.html (published 2026-07-17, captured 2026-08-06, third-party eval, via media)
  15. Juejin (GLM-5.2 AA 51 open-source SOTA), juejin.cn/post/7654122741078671402 (published 2026-06-24, captured 2026-08-06, third-party eval, via community)
  16. 36Kr EU / InfoQ (Anthropic: never advocated banning open weights), eu.36kr.com/zh/p/3914871510914178 (published 2026-07-27, captured 2026-08-06, media report)
  17. GeekPark (Jensen Huang's first X post, 25-company open letter), geekpark.net/news/367973 (published 2026-07-25, captured 2026-08-06, media report)
  18. OSChina / Bloomberg (Chinese open weights spared US safety tests), oschina.net/news/488173 (published 2026-08-05, captured 2026-08-06, media report of Bloomberg reporting)
  19. Juejin / Awakening AI (leaked closed-door meeting notes: 20x compute gap, 12-18 month model gap, 3-6 month target), jxxy.net/ai/articles/liang-wenfeng-deepseek-funding-halt (published 2026-07-26, captured 2026-08-06, media report of unverified leaked notes)
  20. iHeima (GLM-5.2 HLE/GPQA -5%), m.163.com/dy/article/KVFSBKR605118I96.html (published 2026-06-15, captured 2026-08-06, media report)

Disclosure: no sponsors, no affiliate links. Compiled from public sources with capture dates; we did not run any of these models for this piece. The leaked-meeting numbers are flagged as unverified media reports.