China accelerates in AI with cost advantage of up to 100x

What happened?

According to a comprehensive TechRadar analysis of 33 artificial intelligence models from 15 providers, Chinese labs have managed to drastically reduce inference costs—up to 100 times cheaper than Western models—while closing the performance gap. On the GPQA Diamond benchmark, a demanding scientific reasoning test, Western leader Claude Mythos 5 achieves 94.4%, followed by Gemini 3.1 Pro with 94.3% and Claude Fable 5 with 94%. However, Chinese model Qwen 3.7 Max already scores 92.4%, while GLM-5.2 and DeepSeek V4 Pro achieve 91.2% and 90.1% respectively. The difference in scores is minimal (barely 2-4 percentage points), but the cost is abysmal: in some cases, Chinese models cost up to 100 times less per inference token. This is not an isolated phenomenon: since 2023, DeepSeek, Alibaba (Qwen), Baidu (ERNIE) and others have released open-weight models that directly compete with GPT-4, Claude 3 and Gemini, but with prices that challenge the Western cost structure. For example, DeepSeek V2 charged $0.14 per million input tokens versus $10 for GPT-4 Turbo at its launch. The gap has been maintained and widened in 2025.

Why is it important?

This paradigm shift implies that the winner of the AI race will not necessarily be the one with the smartest model, but the one that makes advanced AI affordable for everyone. Historically, Western dominance in AI was based on superior performance and investment capacity. However, the Chinese strategy focuses on infrastructure efficiencies (use of domestic hardware, software optimization), lighter regulation (which reduces compliance costs) and aggressive market strategies (sometimes state-subsidized). Lower inference costs allow small and medium-sized businesses to access capabilities that were previously only available to tech giants. For example, an e-commerce startup can now deploy a chatbot with advanced reasoning for a fraction of the previous cost. Additionally, traffic on platforms like OpenRouter already shows that Chinese models account for nearly half of token volume, while American models have fallen from 74% to 20% in just six months. This data reflects massive adoption in applications where cost is critical. As TechRadar notes: "The winner of the AI race may not be the company with the smartest model, but the ecosystem that makes advanced AI affordable enough to deploy everywhere."

Consequences for the market

The market is splitting into two distinct layers: the mass layer, where 'good enough' has already won, and the frontier layer, where quality matters more. In the mass layer (classification, customer support, content generation), Chinese models are dominating due to cost. For example, logistics companies in Asia already use DeepSeek to optimize routes and reduce operational costs by 30%. In the frontier layer (complex reasoning tasks, advanced coding, autonomous agents), corporate dollars still concentrate on Western models like Anthropic, which captures 40% of enterprise API spending, according to industry estimates. However, the performance gap is narrowing rapidly. On benchmarks like MATH-500 or HumanEval, models like Qwen 3.7 Max already surpass GPT-4o in some tasks. The cost advantage could tip the scales: if a company can achieve 90% accuracy at 1% of the cost, the business decision is clear. It is important to note that part of the price difference is due to a lighter safety model in Chinese open-weight systems. This transfers the responsibility for testing, bias mitigation and regulatory compliance to the companies that deploy them, which can increase indirect costs. In contrast, Western models like Claude include robust built-in guardrails, justifying part of their premium price.

What should readers know?

• The cost advantage of up to 100x is not a mistake: it is the result of infrastructure efficiencies (use of domestic chips like Huawei's, optimization of alternative CUDA kernels), lighter regulation (especially on privacy and bias) and aggressive market strategies (sometimes pricing below cost to gain share).
• For typical commercial applications, Chinese models offer unbeatable value for money, but require greater diligence in security and compliance. Companies must assess privacy risks, biases and potential misuse, especially if they operate in regulated sectors like healthcare or finance.
• The market will not have a single winner: AI models do not have significant network effects (switching from one to another is almost costless, as APIs are interchangeable), and model routing—where an orchestrator chooses the best model for each task—is already standard practice. This favors a diverse ecosystem where expensive high-performance models for critical tasks and cheap models for volume will coexist.
• Historically, we have seen similar dynamics in other technology markets: open-source software (Linux, Apache) displaced proprietary solutions on web servers; Android smartphones swept iOS in global market share. AI could follow an analogous pattern, with the difference that here inference cost is the determining factor.

"The winner of the AI race may not be the company with the smartest model, but the ecosystem that makes advanced AI affordable enough to deploy everywhere."

In summary, we are witnessing a reconfiguration of the AI market. Chinese labs have not only closed the performance gap, but have redefined competition in terms of cost. Companies that adopt a multi-model strategy, combining the best of both worlds, will be better positioned to take advantage of this new era of democratized AI.