Chinese artificial intelligence (AI) developers are undercutting American rivals on token prices, but according to recent analyst reports, many enterprises end up paying more once the tokens needed to finish a task are counted.
The so-called “DeepSeek Moment” in January 2025 wiped nearly US$600 billion off Nvidia’s market value in a single trading day, as investors feared DeepSeek’s low-cost model had exposed American giants’ massive spending on large language models (LLMs) as wasteful. A similar jolt hit markets last month when Moonshot AI released its low-cost Kimi K3 model.
Both episodes cemented a belief across Silicon Valley that Chinese AI held an unbeatable pricing edge. That belief is now unraveling. A growing number of analysts conclude the promise of cheaper Chinese AI is itself a “hallucination,” the very term for a model that confidently states falsehoods.
Meanwhile, the geopolitical divide between the two camps runs far deeper than questions of price and efficiency.
US officials have drafted a letter warning that neutrality is no longer an option in the contest over AI, Reuters reported. Countries that align with Beijing’s competing AI framework could be entirely shut out of the American-led coalition, according to the letter.
In June this year, some 35 nations put their names to a US-drafted AI Opportunity Statement, while a related initiative, Pax Silica, aimed at locking down supply chains for AI models, semiconductors and critical minerals, has since drawn roughly two dozen members, among them Japan, Australia and South Korea. Last month, Chinese President Xi Jinping unveiled a rival bloc, the World Artificial Intelligence Cooperation Organization (WAICO).
Kazakhstan, a critical-mineral-rich country, was caught in between as it signed up to both camps.
“To be part of everything is to be part of nothing. Signature of the Pax Silica Declaration is not merely a membership subscription, but a commitment,” the letter tells Kazakhstan. “It cannot be held alongside membership in duplicative initiatives whose expectations conflict with our own.”
Beijing’s embassy dismissed the ultimatum, saying “such actions will only stifle global AI advances and serve no one’s interests.” The clash shows how the rivalry has expanded past which model performs best, reaching into export controls, cloud access and data-sharing rules that could split the global AI market in two.
“Kazakhstan has become a target of the Trump administration because it holds the critical mineral reserves that advanced technology depends on,” writes a Liaoning-based blogger using the pen name Huhu, in an article titled “The AI Cold War has escalated.” “Washington’s dilemma is that if Kazakhstan is allowed to hedge between both sides, the other 35 signatory countries could demand the same treatment.”
“The US ultimatum has put Kazakhstan in an extremely awkward position,” he says. “Choosing America risks losing China’s market and mineral resources, and choosing China risks being cut out of the US-led AI supply chain.”
That binary choice is not confined to foreign governments. Reuters reported that World Liberty Financial, a cryptocurrency firm 38% owned by the Trump family, was found to be collaborating with WorldClaw, a Hong Kong platform that sells access to AI models. WorldClaw accepts World Liberty’s tokens as payment, earning the family a share of the revenue. The arrangement is not illegal.
Accusations of stolen technology
The US-China rivalry over AI stems from chip export controls Washington has imposed on Beijing since 2019, aimed at slowing China’s access to advanced semiconductors.
The fight intensified after US President Donald Trump returned to the White House in January 2025. At that time, Howard Lutnick, then Trump’s nominee for commerce secretary, accused DeepSeek of skirting those controls to obtain high-end Nvidia chips and to build its models on stolen American technology.
Last month, US Treasury Secretary Scott Bessent said Washington could sanction overseas AI developers caught stealing from American companies. US officials said they found watermarks from US LLMs embedded in several Chinese models, such as Kimi K3.
On paper, the Kimi K3 model charges US$15 per million output tokens, compared with US$25 per million tokens for Anthropic’s Opus 4.8 and US$30 per million tokens for OpenAI’s GPT-5.6 Sol. That makes the Chinese model roughly 40% to 50% cheaper than its American rivals.
However, a study by AlphaSense, a market intelligence platform, found that sticker price is a poor guide to the true cost per solved task. The firm tested 246 financial analysis jobs, including earnings call transcripts, US Securities and Exchange Commission (SEC) filings and acquisition activity.
GPT-5.6 Sol delivered about 20% higher quality than Kimi K3 while costing roughly 13% less on a median basis, the study found. Opus 4.8 scored 13% higher on quality while costing about half as much as the Chinese model when total token usage was factored in.
The report explained why cheaper models often fail on hard tasks. It is not that they are less intelligent, but that they are more likely to search for and rely on the wrong supporting material, what researchers call context, producing a wrong answer as a result. That failure mode does not show up in token pricing at all. In fact, there is a clear price-quality mismatch across all LLMs on the market.
On June 26, Kilo, a workflow platform for AI coding agents, published an Efficient-versus-Frontier-Comparison report. Auto Efficient, its budget routing mode that automatically sends tasks to cheaper models instead of top-tier ones, completed 46.7% of KiloBench task trials at an average cost of just 22 US cents per trial.
The frontier tier, an average of results from Claude Opus 4.8, Claude Sonnet 4.6 and GPT-5.5, completed 65.6% of the same trials at roughly 79 US cents each. Auto Efficient therefore delivered 71% of the frontier average’s completion rate at 72% lower average cost, Kilo found.
“We are still in a period of explosive growth for generative AI, and it has only been a little over six months since Claude Code and Codex launched. No one knows when the market will settle, or what the shape of competition will look like once it does,” a columnist for the Chinese technology outlet TMTPost writes.
“Chinese models have not created any actual impact on the revenue of Anthropic and OpenAI,” he says. “They are not aimed at disrupting anyone, but accelerating agentic coding and workflow tools through cost-effective tokens in China.”
Driven by its leading products and strong demand from enterprise users, Anthropic’s annualized revenue run rate (ARR) climbed from US$9 billion at the end of 2025 to US$47 billion by May 2026.
Anthropic is reportedly planning for an initial public offering (IPO) in October. It was valued at US$965 billion in its last private funding round in May, and could debut above US$1 trillion once shares begin trading.
Chinese observers say Anthropic’s success has pointed to a feasible and sustainable path that Chinese AI firms have yet to follow. Chinese firms have so far focused on building faster, cheaper LLMs rather than serving corporate needs. But they added that Anthropic’s IPO will also widen its lead over Chinese rivals.
Read: US may sanction China’s Moonshot for distilling Anthropic’s Fable
Follow Jeff Pao on X at @jeffpao3
