- Published on
OpenAI and other overseas giants slash prices to challenge Kimi and DeepSeek
- Authors
- Name
- DeepHorus
The artificial intelligence industry is entering a new phase of intense competition. On August 4, OpenAI announced a drastic 80% price cut for its frontier model ChatGPT 5.6 Luna, and a 20% reduction for the mid-tier model 5.6 Terra, directly targeting new-generation models from China's Moonshot AI (Kimi K3) and DeepSeek (V4 Flash). This move marks a further escalation in the Sino-US competition over large AI models, forcing global AI companies to adopt more aggressive pricing strategies to capture market share.
The spark for this price war came from Chinese developers' breakthroughs in cost control. DeepSeek had already shaken the Western AI industry in 2025, and Kimi K3 caused another shock in 2026. Unlike US companies that bet heavily on cutting-edge technology, Chinese teams leverage strong manufacturing and engineering capabilities to create models that are cheaper, leaner in scale, yet almost comparable in high-end capabilities. When new models with similar performance but vastly different prices flooded the market, enterprise users who had long complained about high token costs quickly switched, forcing giants like OpenAI, Google, and Anthropic to rethink their pricing logic.
Looking at the specific numbers, OpenAI's price cuts are staggering. ChatGPT 5.4, released in March, was priced at $2.5 per million input tokens and $15 per million output tokens at the time. Now, GPT 5.6 Luna has dropped to just $0.20 per million input tokens and $1.20 per million output tokensâan 80% reductionâand has entered the competitive range of DeepSeek V4 Pro (input $0.435/output $0.87). The reduced GPT 5.6 Terra (input $2/output $12) is also lower than Kimi K3's pricing (input $3/output $15). Notably, OpenAI did not lower the price of its flagship model; instead, it raised the price of the 5.6 Sol Fast mode to $10 input/$60 output, positioning it against high-end products like Claude Fable 5 with lower latency.
Behind this strategy are deep technical and cost rationales. AI companies can no longer win solely by scaling model parameters and training data; they must reduce per-token costs through inference optimization and architectural improvements. Meanwhile, Nvidia's new Vera Rubin platform claims a 10x improvement in token performance efficiency, providing the hardware foundation for price cuts. However, the problem is that current AI operating and construction costs have not decreased. OpenAI's $300 billion computing resource agreement with Oracle, and its commitment to invest approximately $600 billion in capital expenditures by 2030, make the price-cutting strategy appear contradictory.
The impact of price cuts on the industry is immediate. For enterprise users, the cost of AI applications has dropped significantly, making more scenarios viable, which could stimulate explosive demand growthâthis is exactly the "Jevons paradox" that OpenAI is betting on: more efficient steam engines actually increased total coal consumption. However, for AI companies themselves, squeezed profit margins are inevitable. OpenAI has been suffering losses in its subscription business, losing tens of billions of dollars in 2025. Google's investment in AI infrastructure is nine times its cloud business revenue. Anthropic has achieved annual profitability but relies on limited collaboration with xAI. Analysts worry that price cuts may make the path to covering massive investments with AI business revenue even more distant.
Industry reactions are polarized. OpenAI and Google claim that the price cuts are a move towards "AI for everyone," which will give rise to more application scenarios. Investors, on the other hand, express doubts about profit prospects, with reports that Nvidia may provide OpenAI with $250 billion in investment support to maintain its ability to purchase computing power. Users generally welcome the price cuts, with many developers saying they can finally afford the inference costs of frontier models, though some voices question whether low prices mean compromised service quality.
Looking ahead, as next-generation AI accelerator chips become widespread in data centers by the end of the year, token processing capacity per unit of power could improve another 10-fold. At that point, lower-margin businesses might rely on scale to become profitable. However, this optimistic outlook requires continuous growth in user demand. If the Jevons paradox holds, total AI usage will rise exponentially, and revenue scale could offset the impact of price cuts.
From a macro perspective of the Sino-US industry, this price war reflects a clash between two development models. American companies lead frontier innovation with capital advantages, while Chinese companies reduce costs through engineering innovation. As the capability gap between models narrows, the competitive focus is shifting from "who can build it" to "who can build it at the lowest cost." This contest is not only about corporate profitability but will also determine whether AI technology becomes a profit tool for a few giants or a truly inclusive infrastructure.
