US Lead in the AI Race With China Is Rapidly Narrowing

# US vs China AI Race: How ChatGPT, Gemini, Deepseek, Kimi Agents Compare Source: [https://www.bloomberg.com/graphics/2026-us-china-ai-race/?terminal=true](https://www.bloomberg.com/graphics/2026-us-china-ai-race/?terminal=true) 102030405060ptsSEP\. 2023Best modelBest modelSEP\. 2023 ## US Lead in the AI Race With China Is Rapidly Narrowing ## A spate of compelling releases at budget prices have made China the frontrunner in the race for global adoption\. By[Luz Ding](https://bloomberg.com/authors/AVXV62pP05M/luz-ding)[Spe Chen](https://www.bloomberg.com/authors/AV33fniD9Gc/spe-chen)[Hayley Warren](https://bloomberg.com/authors/ASy8ufuymtE/hayley-warren)August 19, 2026In the battle for global technological mastery being fought in the labs of AI companies, China is rapidly closing the gap with the US\. By several key metrics — usage and cost — it’s even taking the lead\. That’s a growing concern for American firms such as OpenAI and Anthropic PBC, which are chasing trillion\-dollar valuations based on assumptions about their superior capabilities and money\-making prospects\. Despite longstanding efforts by Washington to restrain the speed of development in China through curbs on the sale of US chips needed to power such complex processing, a succession of new models from Chinese firms has shown marked improvements in coding and reasoning\. The latest blockbuster was Moonshot’s Kimi K3, which nearly matched the capabilities of Anthropic’s most advanced model at a fraction of the cost\. #### China’s Kimi K3 Is Closing In on America’s Best AI Multiple independent performance measures rank Moonshot’s new model among top performers 2023202420252026 0204060 Artificial Analysis Intelligence Index⁠Composite benchmark measuring coding, scientific reasoning and general ability - Model author country: - US - China - Other Frontier modelNon\-frontier model Stars mark models that significantly outperform their country’s previous frontier GPT\-3\.5 Turbo o1 preview o1 GPT\-5\.4 Fable 5 Opus 5 Qwen Chat 14B DeepSeek R1 DeepSeek V3\.1 Terminus GLM\-5 Qwen3\.7 Max GLM\-5\.2 Kimi K3 ### Epoch Capabilities Index Metric for models' general capability across 50\+ benchmarks Better →130140150160 ### Vals Index Performance measure across finance, coding and legal tasks Better → Sources:[Artificial Analysis Intelligence Index⁠ v4\.1\.1](https://artificialanalysis.ai/#intelligence),[Epoch AI](https://epoch.ai/eci?view=graph&tab=release-date&subset-view=graph&subset-tab=Software+engineering),[Vals Index](https://www.vals.ai/benchmarks/vals_index) Note: Data as of Aug\. 14, 2026\. Epoch chart only includes models released since 2025\. In order to give a snapshot of the current state of this fierce AI competition,*Bloomberg Businessweek*has pulled together a series of rankings that compare frontier models — the world’s most advanced AI systems — on a range of capabilities from coding to finance to image generation\.*Bloomberg Businessweek*also examined the performance of some of the leading models in a customized vibe\-coding test\. #### Top 10 AI Models Across Specialized Benchmarks - Model author country: - US - China RankModelProviderScore1GPT\-5\.6 Sol \(xhigh\)OpenAI89\.5%2Claude Opus 5 \(max\)Anthropic89\.1%3GPT\-5\.6 Sol \(max\)OpenAI88\.0%3GPT\-5\.6 Terra \(max\)OpenAI88\.0%3Claude Opus 5 \(xhigh\)Anthropic88\.0%6Kimi K3 \(max\)Moonshot85\.0%7Claude Fable 5 \(with fallback\)Anthropic84\.6%7Claude Opus 4\.8 \(max\)Anthropic84\.6%9Grok 4\.5 \(high\)SpaceXAI81\.6%10Qwen 3\.8 MaxAlibaba81\.3%Source:[Artificial Analysis Terminal\-Bench 2\.1 Leaderboard](https://artificialanalysis.ai/evaluations/terminalbench-v2-1) Note:Terminal\-Bench 2\.1 evaluates AI models on practical software tasks, such as fixing bugs, setting up servers and managing files\. Data as of August 12, 2026 With aggressive pricing, Chinese AI offerings are proving irresistible to developers and businesses\. Chinese releases account for the single biggest share of generative model downloads on Hugging Face Inc\., the go\-to hosting platform for AI models\. US companies are among those piling in\. Airbnb Inc\., food\-delivery giant[DoorDash Inc\.](https://x.com/andyfang/status/2074252174226493584?s=46)and top crypto exchange[Coinbase Global Inc\.](https://x.com/brian_armstrong/status/2070670644577280109?s=46)have all adopted Chinese models hosted on local servers\. In August, Alibaba Group Holding Ltd\. disclosed its open\-weight models have accumulated more than[3 billion downloads](https://www.bloomberg.com/news/articles/2026-08-15/alibaba-ai-models-hit-3-billion-downloads-passing-meta-google)in six months, eclipsing Meta Platforms Inc\., Alphabet Inc\. and domestic peers to become the world’s No\. 1 artificial\-intelligence model family\. The increased popularity of Chinese AI models has triggered concern that President Donald Trump’s administration will take action to limit their availability in the US, similar to the American government’s move to slap controls on imports of Chinese electric vehicles and solar panels\. Yet nearly 200 US companies have spoken out against a ban on Chinese AI models — because it would increase their costs and hurt their ability to compete internationally\. What’s more, these models are so\-called open\-weight, which means a model like Kimi K3 can be downloaded, copied and run locally by anyone\. As a result, any ban would be hard to enforce\. Given the rapid evolution of each new release in this race for innovation, which Meta’s Mark Zuckerberg has called likely the most competitive in history, it’s unlikely a single lab will stay ahead for long\. “They are all going to be caught up by others at some point,” said computer science Professor Ruslan Salakhutdinov, who taught Moonshot founder Yang Zhilin during his PhD studies at Carnegie Mellon University\. The narrowing race comes at a crucial time for the US firms\. Anthropic is considering an initial public offering as soon as October, while OpenAI is now looking at going public next year\. Anthropic raised funds at a $965 billion valuation in May\. OpenAI was worth $852 billion based on its most recent funding round in March\. That dwarfs the $35 billion Moonshot is valued at, but then, few outside tech circles had heard of the Chinese company before last month\. DeepSeek — which upended financial markets with the shock release of its R1 reasoning model in January 2025 — is in the middle of a massive fundraising and has begun preparations for an IPO as soon as this year\. Bloomberg in partnership with researchers at Vals AI, an independent AI evaluation and benchmarking platform, tested seven models from frontier Chinese and US companies to see how they performed in a real\-world task\. They were asked to create a fictional coffee e\-commerce site called Brewberg using the same prompts\. Most of the models scored 100% functional accuracy despite occasional design misses, but with very different price tags\. The experiment employed the top performing models in July from Anthropic and all the Chinese firms, as well as more affordable models from OpenAI and Google\. All of the companies have a range of price points and capabilities\. When it comes to price, Chinese firms are dramatically cheaper\. Moonshot charges $15 per million output tokens for Kimi K3, while DeepSeek’s V4\-Pro model costs $3\.96 for 1 million tokens at peak hours – and half that at non\-peak hours\. Anthropic’s state\-of\-the\-art Fable 5 service is priced at $50\. Undercutting the competition is a tried and tested strategy in China, adopted by firms from tech giant Tencent Holdings Ltd\. to Luckin Coffee Inc\. and e\-commerce platform, Temu’s owner PDD Holdings Inc\. to gain market share\. Such is the pervasiveness of price wars in different industries that the strategy has given rise within China to the term “involution,” or destructive competition, which has added to deflationary pressures and spurred repeated warnings from Beijing to corporate executives\. That discount is of growing value to American business owners as their dependence on AI increases\. Ben Cera’s start\-up Polsia relies on an army of AI agents to automate workflows for businesses\. When his platform went viral, his costs went vertical\. To save Polsia from potential bankruptcy, he shifted to models by Chinese firms, including Shanghai\-based MiniMax\. “At one point I was like, I don’t have a choice,” Cera said from San Francisco\. Switching to MiniMax’s M2\.7 hosted on US servers for most of his company’s needs cut his AI spending to $100,000 from $1 million in a single month\. Cera still prefers Anthropic, saying it has stronger ethical guardrails related to customers’ requests and less error\-prone nature\. “If the cost wasn’t an issue, I would just use Anthropic,” he said, saying Chinese models sometimes display Chinese characters without warning\. But he’s figured out a compromise: using Chinese models for night\-time or background tasks that aren’t urgent, while using Anthropic for live functions\. Customers, he said, don’t care about which model is being used\. “They love it when it works, and they hate it when it doesn’t\.” Looking under the hood of our Brewberg test, it’s possible to see how quickly costs rack up when relying on a single, advanced agentic model that runs autonomously without human intervention to steer the process or cap usage\. Anthropic’s Claude Fable 5 AI model charged the most at $48\.99 because of a higher token price combined with many turns of rigorous debugging and refining, which made its final bill much higher\. #### What’s Driving the Higher Claude Fable 5 Bill? Cost per action of agentic workflow in building the Brewberg website Fable 5 began byreviewing the files and confirming the setup,which cost about 34 cents\. Fable 5’s most costly phase wasbrowsing and debugging\.It took screenshots to verify output \(which increased token usage far beyond text input\), returned to the code to fix issues, and repeated that loop until the result met its standard before submitting\. 1/161ExecutionTotal cost:$0\.34 ### China Wants Market Share Most Chinese AI labs have embraced open\-weight software, which allows users to customize, download and, crucially, host the technology on foreign cloud services\. That strategy has been central to attracting users overseas because it sweeps away much of the disquiet about data security — even at the expense of direct returns\. That’s in contrast to top American labs, which have mostly kept their models closed\-weight\. While this approach allows US companies such as Anthropic to have tighter control over their technology, it results in a higher price tag for users who can only access the models through paid APIs or subscriptions\. Chinese labs are more comfortable than their Silicon Valley counterparts with tolerating lower profits, said Kevin Xu, the founder of US\-based tech fund Interconnected Capital\. “Because AI is so new, I think all these open models and companies are still in customer grabbing or land grabbing mode,” he said\. #### Chinese Model Usage Has Surged Past US Peers on OpenRouter Share of monthly tokens, by model author country 020406080%Jan\.2025JulyJan\.2026JulyUSChina Source: OpenRouter Note: OpenRouter data doesn’t capture traffic sent directly to providers and represents only a portion of global AI use\. It’s working — the use of Chinese models overtook US platforms globally for the first time in June, and accounted for more than 60% of market share last month, on OpenRouter, a tech platform that offers software developers access to hundreds of AI models\. It is a widely watched gauge of model usage despite tracking just a fraction of global AI consumption\. The US, parts of Europe and Asia now favor Chinese labs, according to the same data\. On Hugging Face, Chinese AI models account for 41\.4% of generative model downloads among developers, 5 percentage points higher than US models\. Robert Lea, an analyst at Bloomberg Intelligence, said for now none of China’s AI labs have a clear path to profit, despite the rising popularity of their increasingly powerful models\. Lea believes the current business approach, which is over\-reliant on utility\-like token supply at dirt cheap prices, will ensure the over\-supplied sector remains loss\-making for the next three years\. #### Chinese AI Models Have Gained Traction Overseas Top 10 markets of highest token usage volume by company on OpenRouter - Model author country: - US - China 377 trillion tokens 53% 42% market share DeepSeekDeepSeekxiaomixiaomiMiniMaxMiniMaxanthropicanthropicgooglegoogleOpenAIOpenAI tencenttencent 17T 53%45% 14T 64%33% 13T 98% 11T 53%41% 11T 52%44% 11T 62%33% 9T 51%43% Source:[OpenRouter](https://openrouter.ai/rankings) Note: Token volume represents the data from January 2026 to July 2026\. Data excludes direct purchases or subscriptions traffic from model developers and pre\-release anonymous models from other labs issued by OpenRouter\. Commercial use aside, Chinese models are struggling to reach everyday users overseas\. OpenAI continues to reap the benefits of being the developer that made generative AI mainstream with the launch of its ChatGPT app in mid 2023\. Total downloads since 2024 are significantly higher than even those of its US rivals combined, according to data from market intelligence firm Sensor Tower, which excludes downloads in China\. #### US AI Assistant Apps Lead Global Mobile Downloads Monthly downloads from Apple’s App Store and Google Play, excluding China US Apps050100150200MGrokPerplexityClaudeCopilotChatGPTleadsasgenerativeAIgainsglobaltractionGemini'sviralimageeditingfeaturesparksa430%surgeChina Apps020MCici/DolaKimiDeepSeek'sJanuarydebutdraws25\.7Mdownloads202420252026 Source: Sensor Tower Note: Data only include apps with peak monthly downloads above one million\. Data from January 2024 to July 2026 The importance of the industry to China can’t be overstated: not only does the country stand a good chance of dominating a technological sphere with the potential to revolutionize industry and economies around the world, but surging exports of tech hardware are becoming vital to offsetting weak domestic consumption and shoring up faltering economic growth\. Beijing is preparing to spend around 2 trillion yuan \($295 billion\) over the next five years building data centers across the country as part of efforts to boost domestic chipmaking and reduce reliance on US firms like Nvidia Corp\. and Advanced Micro Devices Inc\. China is also ramping up its power generation capacity, much of which is coming from renewables such as wind and solar\. That allows existing and upcoming data centers to run on cheaper, cleaner electricity from a stable national grid\. By contrast, the US faces an energy bottleneck because of surging demand from new computing facilities that are significantly[driving up](https://www.bloomberg.com/news/articles/2026-05-14/data-centers-drive-76-rise-in-power-bills-on-largest-us-grid)power costs and often met with local backlash\. That’s prompting US hyperscalers to build data centers overseas\. Lower labor costs are also boosting China’s hand\. While top talent can get payment packages comparable to their peers in the US, especially those that jumped ship from Anthropic or OpenAI, most researchers in Chinese labs are still paid significantly less\. A post for a[large language model engineer](https://www.zhipin.com/web/geek/jobs?city=101010100&query=%E5%8C%97%E4%BA%AC%E6%B7%B1%E5%BA%A6%E6%B1%82%E7%B4%A2%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD%E5%9F%BA%E7%A1%80%E6%8A%80%E6%9C%AF%E7%A0%94%E7%A9%B6%E6%9C%89%E9%99%90%E5%85%AC%E5%8F%B8https://www.zhipin.com/web/geek/jobs?city=101010100&query=%E5%8C%97%E4%BA%AC%E6%B7%B1%E5%BA%A6%E6%B1%82%E7%B4%A2%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD%E5%9F%BA%E7%A1%80%E6%8A%80%E6%9C%AF%E7%A0%94%E7%A9%B6%E6%9C%89%E9%99%90%E5%85%AC%E5%8F%B8)at DeepSeek in Beijing is offering between 480,000 and 840,000 yuan \($71,230 and $125,650\) per year, whereas a San Francisco[training engineer](https://openai.com/careers/researcher-training-san-francisco/)position at OpenAI is listed for at least $360,000\. None of the AI companies featured provided a comment for this story\. #### China’s Kimi K3 Nears US Claude Fable 5’s Performance At Cheaper Cost Weighted\-average cost of completing model capability benchmark tasks - Model author country: - US - China - Other 0255070 Better performance → 012$3 Cost per taskCheaper→ Opus 5Fable 5Kimi K3Qwen3\.8 MaxMuse Spark 1\.2DeepSeek V4 ProDeepSeek R1 Source: Artificial Analysis Intelligence Index v4\.1\.1 Note: Data as of Aug\. 14, 2026 There is a deepening sense of unease in the US over the rise of China’s AI technology\. A report by US consulting firm Booz Allen Hamilton has also argued that some should be banned for use by the American government and other critical infrastructure because they produced less secure code and have a political bias aligned with the views of the Chinese government\. The US should have a massive computing advantage over China, because of export controls on Silicon Valley’s most powerful chips and the machines that manufacture them\. Despite efforts to reduce reliance on imports, Chinese manufacturers like Huawei Technologies Co\.’s Ascend and Alibaba’s T\-Head are struggling to produce enough to meet surging demand from domestic labs\. Saif Khan, a technology fellow at Washington, DC\-based think tank the Institute for Progress, said US computing power is around 10 times that of China\. But if export controls were perfectly enforced, that would grow ten\-fold\. The latest flurry of releases has also added to suspicions that Chinese labs built their models distilling knowledge from American AI systems\. Anthropic and OpenAI have repeatedly accused Chinese AI labs of this process of distillation, which mimics another model’s behavior but typically lacks its full capacity\. The technique is widely adopted in the industry, but[OpenAI](https://openai.com/policies/row-terms-of-use/)and[Anthropic](https://support.claude.com/en/articles/12326764-can-i-use-my-outputs-to-train-an-ai-model)prohibit companies from using their outputs to train competitive models, and the White House has been voicing concern of intellectual property theft\. US Treasury Secretary Scott Bessent has floated unspecified[sanctions](https://www.bloomberg.com/news/articles/2026-07-21/bessent-says-us-will-scrutinize-chinese-ai-models-for-ip-theft)if model developers are found to have stolen intellectual property, while Trump has signaled that his administration is weighing additional controls\. Su Hang, a computer science researcher at China’s most prestigious STEM school Tsinghua University, dismissed talk that China’s advances are solely reliant on distillation\. “China still has a gap with the US in many aspects of AI, but Chinese developers like Kimi and DeepSeek have amassed the best people,” he said\. “That high concentration of talent has enabled such breakthroughs\.” Yet limiting Chinese competition over national security concerns could backfire, given the economy is increasingly reliant on affordable AI, as Trump acknowledged\. “We have to be careful in both ways,” Trump told reporters July 29\. “We don’t want to restrict them where all of a sudden, we come in second to China\.” Businesses need access to cheap, abundant intelligence, whether it’s Chinese, American or from any other country, said Harry Godfrey, the executive director of the Little Tech Association, which represents nearly 200 US companies using Chinese models\. He has appealed to Trump to identify the security risks and respond to them with “the lightest touch\.” “Our strength is the ability to have a really open, really dynamic market economy with lots of different people without fear of political favor or retribution and just make the best products and compete with one another\.” #### Combined Valuation of Major Chinese AI Labs Equals About Half of Anthropic’s Market valuation of major AI labs in US and China Tech conglomerate with AI lab Anthropic$965BOpenAI$852BMeta$1\.4TGoogle$4\.2T Alibaba$299BZhipu$62BDeepSeek$45BMoonshot$35BMiniMax$14\.2B Sources: Bloomberg; company press releases Note: Valuations for unlisted companies are based on closed funding rounds\. Data as of Aug\. 18, 2026 And any US action could trigger a backlash from Beijing\. China’s Commerce Ministry, which oversees export controls,[warned](https://www.mofcom.gov.cn/syxwfb/art/2026/art_7f1622463a7c48ef9fad600ce0ef702f.html)last month that it would take “all necessary measures” in response to US threats\. Beijing has already demonstrated its ability to[hit back](https://www.bloomberg.com/news/articles/2026-06-22/china-places-two-us-rare-earths-producers-on-export-control-list)at US moves against its tech companies by placing export controls on rare earths\. The leaders of the world’s two largest economies are due to meet next month in Washington, with AI policy set to be a key focus, but experts say neither side is expected to make significant compromises given the intense tech race\. Speaking at the World AI Conference in Shanghai last month, President Xi Jinping urged for an inclusive approach, saying that “AI development should not be a solo performance by a single country, but a symphony of international cooperation\.” The US is caught between businesses looking to save costs in AI and the risk of China winning the race on price, adoption, and, possibly one day, technology\. Flo Crivello, founder of AI assistant Lindy, described the ceaseless race for the most sophisticated AI as overserving the needs of many\. He hasn’t looked back since ditching Anthropic for DeepSeek in June, slashing AI spending by 90%, saving millions of dollars that can be spent on headcount, marketing and an off\-site event in California’s scenic wine region of Napa Valley\. “You do not need God to write your emails,” he said\.