Chinese artificial intelligence models are rapidly narrowing the performance gap with leading Western models while cutting inference costs to as little as one-hundredth, an analysis said. The competitive landscape in the AI market is changing, it said. The assessment said the new battleground is how cheaply and widely companies can supply high-performance AI, rather than a race solely for top performance.

According to IT outlet TechRadar on July 28 local time, the winner in the AI race is increasingly likely to be not the single best model, but the company that can build an ecosystem capable of deploying high-performance AI at low cost and at scale. The analysis compared inference performance, operating costs and deployment methods for 33 models developed by 15 AI providers.

In the GPQA Diamond benchmark for scientific reasoning performance, U.S. companies maintained an edge. Anthropic's Claude Mythos 5 scored the highest at 94.4 percent, followed by Google Gemini 3.1 Pro at 94.3 percent and Claude Fable 5 at 94.0 percent.

Chinese AI models also narrowed the gap significantly. Alibaba's Qwen 3.7 Max scored 92.4 percent, GLM-5.2 scored 91.2 percent and DeepSeek V4 Pro scored 90.1 percent, and were assessed to have entered a competitive range similar to top-tier models.

The difference appeared more in cost than performance. TechRadar analysed that low inference costs go beyond doing the same tasks more cheaply and are changing the scope of companies' AI use.

With the ability to handle more requests on the same budget, AI can be expanded beyond a few core tasks into areas such as customer support, content creation and internal automation, it said. It also assessed that an environment is forming in which startups and mid-sized companies can adopt top-tier AI at relatively low cost.

The market also appears to be splitting into two segments. In mass-market use cases where a certain level of performance is sufficient, such as customer consultation, document writing and content creation, price competitiveness is becoming more important than benchmark scores.

Over the past year, the share of U.S. models in traffic on AI platform OpenRouter fell from about 74 percent to around 20 percent, while the share of Chinese models expanded to nearly half. By token usage, Chinese models were assessed to already hold an advantage.

By contrast, in high-difficulty markets that require complex reasoning, such as coding and AI agents, top-performing models still hold significant value, the analysis said.

For tasks that require multiple steps to be carried out consecutively, even small performance differences can greatly affect the final outcome. For this reason, companies' AI budgets are also continuing to concentrate on top-performing models. TechRadar reported that Anthropic accounts for about 40 percent of enterprise API spending based on demand for coding and AI agents.

Chinese companies' open strategy was also cited as a factor driving adoption. GLM-5.2, DeepSeek V4 Pro and Qwen 3.5 397B are provided as open weights, allowing companies to run them directly on their own servers. This offers advantages including gaining control over data and reducing dependence on a specific AI provider's pricing policy or service changes.

TechRadar assessed that Chinese AI companies are not simply offering low-cost alternatives to Western models, but are presenting a new AI adoption model built on low cost, openness and operational control.

Debate over regulation is also continuing. In a recent open letter, Microsoft Chief Executive Satya Nadella (사티아 나델라) and Nvidia Chief Executive Jensen Huang (젠슨 황), among others, argued that excessive regulation of open AI models could weaken U.S. competitiveness. They stressed that expanding open AI models could lower technology costs and spur innovation across industries.

The spread of Chinese AI models also has political limits. TechRadar pointed out that Chinese models are technically open, but may restrict answers or reflect the Chinese government's position on questions about the Tiananmen Square crackdown, Taiwan, Xinjiang Uyghurs, Hong Kong protests and President Xi Jinping.

AI security firm Enkrypt AI also said that its analysis of DeepSeek R1 found that about 91 percent of sensitive questions related to China maintained pro-China responses. This suggests that even if a company runs a model on its own servers, political bias may not be fully removed.

TechRadar analysed that Western AI companies still hold an edge in top performance, but are finding it increasingly difficult to monetise that at high prices across all markets. It forecast that in the mass market, 'good enough' performance and low cost have already become more important competitiveness, and that future AI competition will be decided by who builds a cheaper and more scalable AI ecosystem rather than by benchmark scores.