Alibaba and DeepSeek collapsed prices for “brains” for AI

While Alibaba was boasting about the largest model in its history, neighboring DeepSeek quietly rolled out a system that is cheaper than literally all competitors in the world — including the top-end Anthropic model, the difference with which is up to a hundred times.

Pennies for intelligence

Let's start with what really blew up the industry, suggests xrust. The research company Artificial Analysis from San Francisco has calculated that running a standard test on the DeepSeek V4-Flash model costs about 3 cents (about 2.5 rubles at the current exchange rate). For comparison, the same test on Kimi K3 from Moonshot AI costs 86 cents (about 69 rubles), on GPT-5.6 Sol from OpenAI — $1.86 (about 150 rubles), and on Claude Fable 5 from Anthropic — already $3.15, that is, more than 250 rubles. The difference is more than a hundred times.

Formally, the price for DeepSeek tokens is also ridiculous: $0.14 per million input tokens and $0.28 per million output tokens—that’s about 11 and 22 rubles, respectively. But it's not just about the price list. Analysts specifically consider not the bare price of a token, but the cost of solving a specific problem as a whole: a model can be cheap “on paper,” but require three times as many steps to get to the answer, and ultimately cost more.

2.4 trillion giant

Meanwhile, Alibaba (9988.HK) on Monday unveiled the Qwen3.8-Max, its biggest model in its history, and the market reacted immediately: the company's shares jumped 7% on the Hong Kong exchange. The model has 2.4 trillion parameters — these are the very numerical coefficients on which the system “learns” to recognize patterns and generate answers. Its direct competitor has slightly more parameters: Kimi K3 from Moonshot AI, released a month earlier, accelerated to 2.8 trillion.

The number of parameters itself does not mean anything in terms of quality — a model with a larger number of coefficients is not necessarily smarter. But it is precisely this indicator that has become the unspoken currency with which Chinese companies measure themselves against each other and with Western developers, demonstrating the scale of the invested computing power.

On the Arena.AI platform, where users compare models blindly, Qwen3.8-Max immediately became the best Chinese model among text systems — although it is still inferior to Claude Fable 5 and three versions of Opus from Anthropic. In the category of image and video processing, Qwen3.8-Max took second place in the world, behind only one of the variants of Fable 5.

Both models — Qwen3.8-Max and Kimi K3 — can work with text, pictures and videos simultaneously and are able to “digest” up to a million tokens at a time: these are hundreds of pages of documents, voluminous program code or a weighty legal agreement, downloaded in their entirety. Alibaba plans to release Qwen3.8-Max to a wider audience next week. According to the company, the model has already managed to independently complete a software development project in 16 days — thanks to the “mixture of experts” architecture, where not the entire system is connected to a specific request, but only part of it: about 95 billion parameters out of 2.4 trillion are simultaneously used, which keeps computation costs in check.

Why does China need open weights

Both models have one more thing in common — open weights. Unlike OpenAI, Anthropic and Google, which keep the internal structure of their systems secret, Alibaba and DeepSeek make the basic parameters publicly available, allowing developers around the world to download and modify the models to their liking.

As explained by the chief analyst of the Omdia research company Lian Jie Su, Chinese companies have found a niche: not every business task requires the most powerful model on the market — it is much more important that the system is good enough, transparent and affordable, and it is precisely this request that open-weight models fill.

Behind the seemingly modest prices of DeepSeek there is a completely pragmatic calculation: the company is preparing for a potential IPO, and it has already caused a sensation — at the beginning of 2025, the R1 and V3 models collapsed shares of technology giants around the world and made investors wonder whether American companies were overpaying for the creation AI. Apparently, DeepSeek plans to repeat the effect of surprise not with a loud announcement, but with a bare price list.

Sources:

• https://www.scmp.com/tech/article/3362738/alibabas-ai-model-qwen38-max-made-widely-accessible-ahead-open-weights-release
• https://www.technology.org/2026/08/03/deepseek-v4-flash-cheapest-ai-model-to-run/

https://artificialanalysis.ai/models/deepseek-v4-flash

Xrust Alibaba and DeepSeek collapsed prices for “brains” for AI