August 5 News.AIIndustry has entered a new stage of development. To deal with itChinaCompetition for a new generation of models (e.g. Kimi K3 in the dark side of the moon and V4 Flash in DeepSeek) has led to significant reductions in product prices by major AI firms worldwide, while increasing the capacity of entry-level models。
Recently this strategy was taken by OpenAI. It reduced the price of the base front model ChatGPT 5.6 Luna by 80% (per million Token) and lowered the price of the middle model 5.6 Terra by 20%. And just over a week ago, Google just launched the lower-priced Gemini 3.6 Flash and 3.5 Flash-Lite models. Anthropic, although there was no direct price reduction, replaced the lowest-priced Opus 4.8 model with a more capable Claude 5.0 and maintained the same price。
Increased global competition is driving AI capacity to become cheaper, but it may also be at the expense of the profitability of these large companies. In previous months, a number of large AI companies had announced cost cuts and limited the use of AI technology, even those companies that had promoted AI under Elon Mask’s banner. Although more and more employees are using AI, current reports indicate that real productivity gains have not met expectations。
Further competition
THE DIFFERENCES BETWEEN CHINA AND THE UNITED STATES OF AMERICA ' S AI DEVELOPMENT PATH REFLECT TO SOME EXTENT THEIR LONG-STANDING INDUSTRIAL ADVANTAGE。
USABUSINESSES TEND TO INVEST HUGE AMOUNTS OF MONEY IN CUTTING-EDGE TECHNOLOGICAL DEVELOPMENT, WHILE CHINESE DEVELOPERS RELY ON STRONG MANUFACTURING AND ENGINEERING CAPABILITIES TO CREATE A CHEAPER, LEANER-SIZED, BUT ALMOST SUBLIME AI MODEL OF HIGH-END CAPABILITIES。
DeepSeek had a huge impact on Western AI developers in 2025, and Kimi K3 again had a similar impact in 2026. Individually, these events are enough to put pressure on companies like OpenAI, Google and Anthropic。
However, in the previous months, many AI-intensive businesses had complained about the rapid increase in Token costs, and the whole industry had been strongly shaken when new models of market proximity, but significantly lower prices, had emerged。
Today, large AI companies cannot simply compete by increasing the number of model parameters and training data. They have to really start the price competition, and OpenAI has significantly reduced the price of some of the models。
It's worth noting that OpenAI did not lower the price of the most powerful flagship model. In fact, the price of the faster version of the model has actually risen. But its front-level model is now at its lowest ever, and AI's ability to increase and price decline is alarming。
In March of this year, OpenAI released ChatGPT 5.4, a model with a stronger intelligence capability, at a price of US$ 2.5 per million entered by Token (note: the current exchange rate is about RMB 16.9) and US$ 15 per million exported (the current exchange rate is about RMB 101.5)。
Today, the GPT 5.6 Luna price has fallen to US$ 0.20 per million input to Token (current exchange rate is about RMB 1.4) and US$ 1.20 per million output (current exchange rate is about RMB 8.1)。
This means that a front-line model has gone down significantly from publication to price, and has been going through less than four months。
As a result of this price reduction, Luna ' s prices have entered the competition of DeepSeek V4. The price of the DeepSeek V4 professional edition is US$ 0.435 per million input token (current exchange rate is around 2.9 yuan), US$ 0.87 per million output token (current exchange rate about 5.9 yuan)。
GPT 5.6 Terra is a more capable model, but after a reduction in the price of 20%, its price is also reduced to US$ 2 per million input (current exchange rate is about RMB 13.5) and US$ 12 per million output (current exchange rate is about RMB 81.2), which is lower than Kimi K3 per million input to RMB 3 per million input (current exchange rate is about RMB 20.3) and token 15 per million output (current exchange rate is about RMB 101.5)。
At the same time, GPT 5.6 Sol maintains the prices of Token (current exchange rate approximately RMB 33.8) and 30 (current exchange rate approximately RMB 203) per million. OpenAI even raised the price of its top model: 5.6 Sol ' s Fast Model charge is US$ 10 per million input of Token (current exchange rate approximately HK$ 67.7), output of US$ 60 (current exchange rate approximately HK$ 406) to provide the same level of intelligence with lower delay and direct competition for flagship front models such as Claude Fable 5 and Mythos 5。
But the question is, do these price-cutting strategies really make a profit for OpenAI
Cost pressure is approaching
Earlier this year, following the announcement by OpenAI of the de facto abandonment of its own data centre for the first party, the lease contract with the “Neoclouds” became even more important。
For example, the agreement with Oracle for the procurement of computing resources of up to $30 billion (the current exchange rate is approximately RMB 2.03 trillion) has become an important basis for the sustainability of OpenAI services。
However, this is all the more obvious now if there has been a prior doubt as to how OpenAI would bear such a cost。
OpenAI is currently running a loss on subscriptions and did not meet key revenue targets earlier this year. And in 2025, companies lost tens of billions of dollars, even if revenues continued to grow。
OpenAI has committed about $60 billion by 2030 (the current exchange rate is about 406 trillion yuan) to calculate resources。
Even if income continues to grow, it may be far from covering such a large cost. Yet today ' s most popular and lowest-priced model reduces prices further, implying that the company ' s profit space may shrink significantly or even disappear completely。
This may also be one of the reasons for the discussion about the possible investment support to OpenAI of $250 billion (at a current exchange rate of about RMB 1.69 trillion)。
Of course, OpenAI is not the only enterprise facing this problem. Over the past year, Google ' s investment in AI infrastructure has been about nine times its cloud business income; Anthropic has recently been able to make an annualized profit and has largely benefited from limited cooperation with XAI in a low-cost rental of the Colossus data centre。
WHILE AI DID NOT SUDDENLY FALL IN OPERATING AND BUILDING COSTS, BUSINESSES WERE LOWERING PRICES AND PROVIDING FASTER AND STRONGER MODELS AT LOWER COST。
On the surface, these figures do not seem reasonable。
"Jevens paradox."
The industry often refers to Jevons Paradox to explain the increase in the spread of AI。
The more efficient coal engines did not reduce coal consumption in the age of Evans, but rather increased total demand for coal as a result of increased use。
TODAY, AI COMPANIES BELIEVE THAT AS AI USAGE COSTS FALL, PEOPLE WILL FIND MORE APPLICATIONS, THUS CONTRIBUTING TO A SIGNIFICANT INCREASE IN OVERALL AI USAGE。
This may be the logic behind the current Token price reduction strategy。
If Token is cheap enough, users will increase their use significantly, thus generating higher incomes through scale growth. And considering that after the end of this year, a new generation of AI-accelerated chips will become more widely available in the data centre, and that the capacity of Token processing in units of power may be 10 times higher, so that businesses with lower profit margins can also profit on a scale。
In addition, the future Vera Rubin platform in Inweida deserves attention. Ying Weidar claims that the platform will lead to 10 times more efficient Token。
In theory, the advantages of the Blackwell and Vera Rubin platforms could make the AI reasoning server a more profitable industry。
But even so, it is hard to imagine that large AI companies can rely on AI business income to cover the huge costs of current inputs, especially when competition is increasing and Token prices are falling。