Can businesses afford the AI they trust?

On 12 October, Meta announced the launch of its next-generation AI model, Muse Wonder, claiming it has achieved world-class performance in multiple tasks. This is a significant achievement after Meta invested heavily in recruiting talent, expanding its data centre, and restructuring its artificial intelligence (AI) team. But does the market truly need another most powerful AI model? This question is gradually becoming a new challenge for the AI ​​industry.

Why pay more to execute the same task?

Over the past month, AI companies worldwide have almost taken turns announcing technological breakthroughs. OpenAI launched GPT-6 Astra, emphasising its capabilities in programming, scientific research and professional work; Anthropic launched Claude Fable 5.1 and Mythos 5.1, claiming top-tier programming and knowledge processing capabilities; Google also joined the competition with Gemini 4 Argon. Each company claims to be at the forefront of technology, creating a dazzling array of claims.

However, the trend of enterprises using AI stands in stark contrast to the hype from these tech companies.

According to statistics from the OpenRouter platform, in Q1 2025, the top ten AI models by quality accounted for 55% of platform usage, but this figure dropped to 14% by Q3 2026. While this data does not represent the entire AI market, it reveals an important message: increasingly more users are abandoning the blind pursuit of the most powerful models and opting instead for lower-priced, but still capable, products.

The reason is not hard to understand. Most of the tasks businesses perform daily using AI, including customer service Q&A, document summaries, email sorting, meeting minutes and data retrieval, do not require state-of-the-art inference capabilities. If a high-level model costs 10 RMB (US$1.50) to process a task while a smaller model costs only 1 RMB, and the results are similar, businesses will naturally choose the latter. When daily usage reaches tens or even hundreds of thousands of times, the cost difference becomes considerable.

This also illustrates that the AI ​​industry is undergoing a significant transformation. In the past, model capability was the main competitive indicator; now, price, speed, reliability and practical benefits are gradually becoming important factors in business procurement. The most powerful model is not necessarily the most suitable model, and technological superiority does not guarantee commercial success.

Multiple models at hand

Therefore, model routing has gained significant attention. Enterprises can establish multi-model platforms, where the system automatically selects the appropriate model based on task complexity, data sensitivity and budget. Simple administrative tasks can be handled by smaller models, complex legal analyses or scientific research by higher-level models, and confidential data can be processed by internally deployed models. This approach balances cost, efficiency and security.

This shift warrants serious consideration from businesses, governments and educational institutions worldwide. Recently, many organisations have rushed to purchase expensive GPUs and build their own AI servers, neglecting the costs of equipment depreciation, power consumption, system maintenance and specialised talent. GPUs are updated extremely quickly; systems built with huge sums of money today may be obsolete in a few years. Even more worrying is the possibility that after purchasing the equipment, it may lack sufficient application scenarios, or the system’s performance may become inadequate as the number of users increases.

A more reasonable approach is for institutions in the same field to collaborate on building shared AI platforms. For example, universities, hospitals or financial institutions could jointly develop domain-specific models, pooling computing power and talent, and then allocating resources according to different work needs. Governments should also encourage such collaborations to avoid redundant investment and wasted resources.

Of course, the most advanced models still have value. Drug development, mathematical proofs, complex programming and scientific exploration require continuous breakthroughs in AI capabilities. However, these high-level needs should inherently differ from the usage scenarios of general administrative work.

Balancing cost, performance and geopolitical security

Businesses also face an AI dilemma between affordability and geopolitical security. As competition between the US and China intensifies, enterprises face a difficult question when choosing AI models: balancing cost, performance and geopolitical security.

Chinese AI companies, such as DeepSeek and Alibaba, offer powerful open-weight models that businesses can download, customise and deploy locally at relatively low cost. These models provide flexibility, reduce vendor dependence, and make AI adoption more affordable, particularly for smaller enterprises.

However, concerns about cybersecurity, regulatory compliance, intellectual property and exposure to China’s technology supply chain discourage many companies from adopting Chinese models. Even local deployment cannot completely eliminate concerns about model provenance, embedded biases and future geopolitical restrictions.

American AI providers, including OpenAI, Anthropic and Google, offer highly capable models and mature commercial ecosystems. Yet their leading models remain largely proprietary, with substantial subscription and API costs. Enterprises must also accept limited transparency, vendor lock-in, and dependence on pricing and policies beyond their control.

This creates a striking paradox. Chinese AI models are increasingly open and affordable but politically sensitive, while American models are commercially trusted by many enterprises yet expensive and restrictive. Companies seeking technological independence may find themselves dependent on one ecosystem or the other.

The dilemma extends beyond technology procurement. AI has become a strategic decision involving national security, corporate governance, operational resilience and long-term competitiveness. Choosing the cheapest model may introduce geopolitical risks, while choosing the safest perceived option may impose unsustainable costs.

A diversified, multi-model strategy could reduce these dependencies, but it also introduces integration and management complexity.

Ultimately, enterprises must confront an uncomfortable question: can they afford the AI they trust, or trust the AI they can afford?

Tech giants are still vying for global leadership, but businesses have already begun to carefully calculate their costs. Ultimately, the value of AI will be determined by actual results. The smartest model may not win the market. Those who produce the best results at reasonable cost are likely to be the ultimate winners.

Related: The real threat of AI: Human error and the loss of critical thought | Why Chinese celebrate AI but Americans hesitate