Anthropic Calls for an AI Pause, Beijing Refuses: Token Costs

👁 112 reads

Dario Amodei, CEO of Anthropic, published a post in September urging the industry to slow down the pace of frontier model development. The argument sounded sensible: a pause would buy time to work on safety. The catch appeared in a single sentence he penned himself - such a pause “would require cooperation with China, an autocratic country with by far the most advanced AI capabilities.” He then outlined how he would like to slow China down: banning sales of advanced chips and semiconductor manufacturing equipment, cracking down on chip smuggling, and restricting remote access to data centers outside China.

Beijing responded within 24 hours. That sets the proper frame for this story: it is not a debate over risk philosophy, but a dispute over who has the right to step on the brakes in someone else’s car.

Chinese Foreign Ministry calls it “fearmongering” while state media cites Cold War playbook

A day after the post went live, state-run outlet Global Times published an editorial titled “Targeting China’s AI: U.S. ‘tech right’ unfolds Cold War playbook”. The argument was direct: proposed restrictions on chips and Chinese AI companies serve to maintain American technological dominance rather than safety. On September 11, Chinese Foreign Ministry spokesperson Guo Jiakun went further, describing Amodei’s appeal as “fearmongering, confrontation, and unfair competition”, adding that such an approach undermines global AI governance.

Amodei gained unusual alignment among domestic competitors - Sam Altman of OpenAI and Elon Musk of xAI leaned in favor of the idea. On the opposite side stood Mark Zuckerberg and Donald Trump, both advocating for rapid, unconstrained development. Even within the United States there is no unified front, meaning a proposal premised on international coordination lacks consensus within its home country.

“I see no reason why China would cooperate with [Amodei’s proposal] and allow itself to be slowed down more than it already is” - Alvin Wang Graylin, senior fellow on Chinese technology at the Center for China Analysis at the Asia Society Policy Institute, speaking to Rest of World

Graylin added a crucial nuance: this does not mean China will never agree to regulate the pace of AI. It simply means it will not agree to terms dictated by an American CEO on his own company blog.

The cost of a million tokens shows no sign of slowing down

Public statements are cheap; API pricing is not. Comparing public API rates across both sides reveals whether anyone is actually easing off the accelerator. A genuine slowdown would mean fewer model releases and stabilizing prices. The exact opposite is taking place: rates are tumbling, which signals surplus capacity and fierce competition for market share.

One million output tokens on DeepSeek currently costs around $1.68. The same million tokens in Anthropic’s top-tier model class costs $75. That is a 44.6-fold difference in unit price for the exact same volume. In practice, a company generating 50 million output tokens a month pays roughly $84 with the Chinese provider versus $3,750 with a top-tier US provider. Over an entire year, that gap amounts to roughly $44,000 on a single workload.

Release velocity tells a similar story. Over the past month, Chinese labs - including DeepSeek, Moonshot with its Kimi model, Alibaba with the Qwen lineup, and Zhipu - rolled out or updated several model variants, many available with open weights for download. None of these launches were delayed or scaled back due to calls for a pause. The numbers make it obvious: the brakes have not been touched anywhere in the industry, and the only metric dropping is pricing.

China’s AI safety framework is closer to Western concerns than it looks

The most compelling part of this story lies beyond politics. On September 14, the Chinese body responsible for cybersecurity standards published the third edition of its Safety Governance Framework. The document warns that AI can autonomously learn, optimize, and undergo “recursive self-improvement” at a pace or in directions that exceed human expectations and control.

This matches the precise register of concerns voiced by safety researchers in Western labs - including former Anthropic researcher Jacob Coxon, who warned of similar trajectories. Beijing rejects Amodei’s diplomatic proposal, but it does not reject his underlying risk analysis.

Zilan Qian, a researcher at think tank GovAI, framed the situation clearly in an interview with Rest of World: she does not believe China and the US view AI risks fundamentally differently, but socioeconomic conditions and political institutions cause different risks to stand out and different voices to be amplified in each nation. In China, there is an added military dimension: Minister of State Security Chen Yixin noted that “modern warfare has entered an era increasingly shaped by artificial intelligence.” Negotiating a pause is inherently difficult when your national security establishment considers the technology central to future defense.

What this means when choosing your AI model provider

If you run a company implementing AI for customer support, document processing, or content generation, the US-China race directly affects your operating budget and your regulatory exposure. On one hand, Chinese open-weight models offer unit economics that are dozens of times cheaper and can be hosted entirely on your own infrastructure. On the other hand, American enterprise APIs provide polished documentation, dedicated commercial support, and established compliance roadmaps for regulations like the EU AI Act.

The operational framework breaks down logically. If you are building a system classified as high-risk under the EU AI Act - such as hiring tools, credit scoring, employee performance analytics, or access to public services - you must thoroughly document model lineage, training data sources, and governance mechanisms. With an open-weight model sourced from China, that compliance burden rests entirely on you as the deployer, as the original vendor will not supply the compliance dossiers required by regulators. It is not an outright ban, but a substantial shift of legal overhead onto your internal team.

Because of this, an effective architectural split emerges: run internal, low-risk, high-volume workloads on efficient open models hosted locally or through low-cost endpoints, where the 44-fold price gap yields tens of thousands of dollars in annual savings. For high-risk, customer-facing pipelines handling sensitive personal data, turn to enterprise providers that supply end-to-end regulatory documentation, even if their API usage costs significantly more.

Hardware break-even points are equally worth calculating. Renting a cloud server with a modern data-center accelerator typically costs roughly $1,000 to $1,500 per month. Running an open-source model in-house only beats cheap Chinese API rates at volumes exceeding roughly 1.5 billion output tokens per month. Against premium US APIs, however, dedicated hosting becomes cheaper at just 20 million tokens per month. That threshold is worth reviewing in your billing logs before signing your next enterprise contract.

A meeting between Donald Trump and Xi Jinping in Washington, where the US side sought to address shared risks from AI, did not change this economic reality. As long as a million tokens costs under $2 on one side and $75 on the other, appeals for a pause will struggle to move past the CFO’s desk. Beijing’s true answer to Anthropic’s proposal was never the foreign ministry’s statement - it was the API rate card.

Fact-checked on October 7, 2026. API pricing based on publicly available provider rate cards on the date of review. Sources: Rest of World, “Dario Amodei wants to slow AI. China isn’t taking orders” (quotes from Alvin Wang Graylin and Zilan Qian via the same report); Global Times editorial from September 11, 2026; third edition of the Chinese Safety Governance Framework released September 14, 2026; EU AI Act regulatory documentation via the European Commission.

// read moreEuropean ChatGPT Alternative Uses Chinese AI Model by DefaultState of AIRestaurant AI Phone Assistant Costs: $0.51 per Call?MagazineDaniela Amodei: Anthropic President and Dario’s SisterMagazine

Kinetyra