Mythos vs DeepSeek: Why cheap AI could tip the US-China tech race

China and the US both witnessed milestone events in the AI sector in April this year.

On 7 April, Anthropic released Claude Mythos Preview, marking a historic breakthrough in cybersecurity. What made Claude Mythos so striking to the industry was its shift of AI cyber capabilities from an “assistive tool” to an autonomous offensive and defensive agent — a step-change qualitative transformation, with general-purpose capabilities emerging as a natural extension of its design and scaling.

Claude Mythos: top performance

Claude Mythos had no formal security training, yet it reached “top-tier hacker” level in vulnerability discovery and exploitation. In terms of time, efficiency and cost, it demonstrated a crushing advantage over traditional cybersecurity paradigms.

First, time taken is compressed. Where elite human teams typically require weeks to months to uncover a complex vulnerability, Mythos can do so overnight, and has even identified a 23-year-old Linux kernel vulnerability within just 90 minutes. 

Next, cost is compressed. The cost for humans to discover a high-risk zero-day vulnerability is generally over US$100,000 per exploit, whereas Mythos operates at under US$50 per run, with batch vulnerability discovery costing less than US$20,000 in total, exponentially lowering the barrier to entry.

Finally, there is a depth of automation. Claude Mythos demonstrates an understanding of logic that goes beyond traditional tools. Conventional fuzzers and scanners rely on brute-force enumeration and are unable to grasp underlying logic, resulting in a high false-negative rate. Meanwhile, Mythos grasps code logic, data flow and control flow, allowing it to precisely identify “logic-based vulnerabilities” that are difficult even for humans to detect.

On the day Claude Mythos was released, the US Treasury Secretary and the Federal Reserve Chair urgently convened a closed-door meeting with the CEOs of major Wall Street banks, classifying it as a systemic financial risk. Central banks and financial regulators in the US, UK and Canada issued coordinated warnings, while major tech giants jointly initiated defensive measures, and the model was not made available to the public.

DeepSeek: a winner in cost

Later that same month, DeepSeek, a leading representative of China’s AI sector, also released its latest version. While the update did not attract significant global attention for its technical capabilities, I think that its impact on China-US AI competition may prove more far-reaching.

Over the past year, DeepSeek’s objective has been to replace Nvidia’s chips with those domestically produced in China. It is well known that US tech giant Nvidia has long been the dominant force in large language model chips. Amid comprehensive China-US competition, the US has sought to curb China’s AI development by restricting its access to Nvidia’s high-end chips.

In this update, DeepSeek’s core advantage — apart from domestic self-reliance — lies in a milestone improvement in cost. The premium version of the Huawei Ascend 950PR is priced at 70,000 RMB (US$10,300), just 28% of Nvidia’s H200 (250,000 RMB) and 47% of the H20 (150,000 RMB). The bulk procurement prices of Hygon’s DCU Shensuan 4th generation and Cambricon’s Siyuan 590 are also only 40-60% of comparable Nvidia products.

Due to their rapid technological iteration, Nvidia chips typically have a depreciation cycle of two to three years. By contrast, domestically produced chips, deeply integrated with DeepSeek’s models, offer more stable technical compatibility, allowing their depreciation cycle to be extended to four to five years.

This implies that the annual depreciation cost of domestic chips is only 50-60% of Nvidia’s, further diluting the fixed cost per token. Based on estimates across the value chain, the cost per token for Chinese AI could fall to one-tenth — or even one-hundredth — of that in the US. 

The deep integration between DeepSeek and domestic chipmakers such as Huawei, Hygon and Cambricon has created a positive feedback loop: model optimisation → improved chip performance → lower costs → broader applications → more data → further model optimisation.

China catching up to US

Many are impressed by Claude Mythos and still see US AI as significantly ahead of China, but this overestimates the advantage of US AI.

The AI industry is highly dependent on the scale of real-world applications, as the continuous stream of data generated by these applications forms the fundamental raw material for AI iteration and development. The cost of tokens is a core expense in industrial applications, much like the cost of steel and coal during the early stages of the Industrial Revolution.

Imagine that, in the early stages of the Industrial Revolution, one country’s coal and steel costs were a tenth of another’s — what would that imply? It would mean its pace of industrialisation could be several times, or even exponentially faster. 

Thus, I am convinced that, as in AI video generation, Claude Mythos’s advantage in the field of cybersecurity will likewise be matched or even surpassed by China’s accelerated catch-up within six months to a year.

Related: When cost and practical application takes priority: China surpasses US in AI adoption | How China’s all-in-one platforms are reshaping the AI battlefield