The artificial intelligence industry has entered a new stage of development. To counter competition from a new generation of Chinese models, such as Moonshot AI’s Kimi K3 and DeepSeek’s V4 Flash, major global AI companies are substantially lowering product prices while improving the capabilities of entry-level models.
OpenAI is the latest company to adopt this strategy. It has cut the price of its foundational frontier model ChatGPT 5.6 Luna by 80% (calculated per million Tokens), while lowering the price of its mid-range 5.6 Terra model by 20%. Just over a week earlier, Google had launched the lower-priced Gemini 3.6 Flash and 3.5 Flash-Lite models. Anthropic has not directly cut prices, but replaced its lowest-priced Opus 4.8 model with the more capable Claude 5.0 while keeping the same price.
Intensifying global competition is making AI capabilities more affordable, but this may come at the expense of the profit margins of these major companies. In previous months, several major AI companies had already announced cost-cutting measures and restricted the scope of AI technology usage, including Elon Musk’s xAI, which had previously strongly promoted AI development. Although more and more employees are beginning to use AI, current reports indicate that actual productivity gains have not met expectations.
The differences between the AI development paths of China and the United States reflect, to some extent, the two countries’ long-standing industrial advantages.
U.S. companies tend to invest huge sums in advancing cutting-edge technology, while Chinese developers rely on strong manufacturing and engineering capabilities to build AI models that cost less, operate at a smaller scale, and are nearly as capable in high-end applications.
DeepSeek delivered a major shock to Western AI developers in 2025, while Kimi K3 had a similar impact in 2026. Taken individually, these events were already enough to put pressure on companies such as OpenAI, Google, and Anthropic.
However, in the preceding months, many companies that use AI extensively had been complaining about rapidly rising Token costs. As a result, the emergence of new models with similar performance but substantially lower prices has sent shockwaves through the entire industry.
Large AI companies can no longer compete simply by increasing the number of model parameters and the scale of their training data. They must genuinely begin competing on price, and OpenAI has already sharply reduced the prices of some models.
Notably, OpenAI has not lowered the price of its most powerful flagship model. In fact, the price of its faster version has increased. However, its frontier-level models have now reached their lowest prices in history, while the pace of improvements in AI capabilities and price reductions is astonishing.
In March this year, OpenAI released ChatGPT 5.4, which featured more powerful agent capabilities. At the time, it was priced at $2.5 per million input Tokens (note: approximately 16.9 yuan at the current exchange rate) and $15 per million output Tokens (approximately 101.5 yuan at the current exchange rate).
Today, GPT 5.6 Luna is priced at $0.20 per million input Tokens (approximately 1.4 yuan at the current exchange rate) and $1.20 per million output Tokens (approximately 8.1 yuan at the current exchange rate).
This means that it took less than 4 months for a frontier model to undergo a substantial price reduction after its release.
Following the price cut, Luna’s pricing has entered the competitive range of DeepSeek V4. The professional version of DeepSeek V4 is priced at $0.435 per million input Tokens (approximately 2.9 yuan at the current exchange rate) and $0.87 per million output Tokens (approximately 5.9 yuan at the current exchange rate).
GPT 5.6 Terra is the more capable model, but after a 20% price cut, its price has also fallen to $2 per million input Tokens (approximately 13.5 yuan at the current exchange rate) and $12 per million output Tokens (approximately 81.2 yuan at the current exchange rate), below the widely watched Kimi K3 ($3 per million input Tokens, approximately 20.3 yuan at the current exchange rate, and $15 per million output Tokens, approximately 101.5 yuan at the current exchange rate).
Meanwhile, GPT 5.6 Sol remains priced at $5 per million input Tokens (approximately 33.8 yuan at the current exchange rate) and $30 per million output Tokens (approximately 203 yuan at the current exchange rate). OpenAI has even raised the price of its top-tier model: the Fast mode of 5.6 Sol costs $10 per million input Tokens (approximately 67.7 yuan at the current exchange rate) and $60 per million output Tokens (approximately 406 yuan at the current exchange rate). It provides the same level of intelligence with lower latency, directly competing with flagship frontier models such as Claude Fable 5 and Mythos 5.
But the question is: Can these pricing strategies really generate profits for OpenAI?
Earlier this year, after OpenAI effectively abandoned plans to own first-party data centers, its leasing contracts with “new cloud computing providers” (Neoclouds) became even more important.
For example, its computing-resource procurement agreement with Oracle, worth as much as $300 billion (approximately 2.03 trillion yuan at the current exchange rate), has become an important foundation for the continued operation of OpenAI’s services.
However, if observers had previously questioned how OpenAI could bear such costs, those doubts have now become even more pronounced.
OpenAI is already continuing to lose money on its subscription business and failed to meet key revenue targets earlier this year. In 2025, even as revenue continued to grow, the company still lost tens of billions of dollars.
OpenAI has committed to investing approximately $600 billion (approximately 4.06 trillion yuan at the current exchange rate) in computing resources by 2030.
Even if revenue continues to grow, it may still fall far short of covering such enormous costs. Now that the most popular and lowest-priced models have been made even cheaper, the company’s profit margins may shrink substantially or disappear entirely.
This may be one reason behind reports that Nvidia could provide OpenAI with $250 billion (approximately 1.69 trillion yuan at the current exchange rate) in investment support.
Of course, OpenAI is not the only company facing this problem. Over the past year, Google’s investment in AI infrastructure was approximately 9 times its cloud-business revenue; Anthropic’s recent ability to achieve annualized revenue profitability was also primarily attributable to a limited, low-cost arrangement with xAI to lease the Colossus data center.
The costs of operating and building AI have not suddenly declined, yet companies are continuously lowering prices and providing faster, more powerful models at lower costs.
On the surface, these figures do not seem to make sense.
The AI industry often cites the “Jevons Paradox” to explain why AI adoption is accelerating.
In Jevons’s era, more efficient coal-powered engines did not reduce coal consumption. Instead, broader usage led to an increase in total demand for coal.
Today, AI companies believe that as the cost of using AI falls, people will discover more use cases, driving a substantial increase in overall AI usage.
This may be the logic behind the current strategy of cutting Token prices.
If Tokens become cheap enough, users will significantly increase their usage, generating higher revenue through scale. In addition, as next-generation AI accelerator chips become more widespread in data centers after the end of this year, Token-processing capacity per unit of power may increase 10-fold, allowing businesses with lower profit margins to achieve profitability through scale.
Nvidia’s future Vera Rubin platform is also worth watching. Nvidia claims that the platform will further deliver a 10-fold improvement in Token performance efficiency.
In theory, the advantages of the Blackwell and Vera Rubin platforms could make AI inference servers a more profitable industry.
Even so, it is difficult to imagine that major AI companies will be able to cover their current enormous costs solely through AI-business revenue, especially as competition intensifies and Token prices continue to decline.