Just hours after OpenAI announced GPT-5.5, DeepSeek has published the long-awaited, completely new version of its flagship model, DeepSeek V4. The timing of this information's publication is probably not accidental. The narrative surrounding the new models of Western giants revolves around, among other things, the rising costs of using them, and the new DeepSeek (like its predecessor) promises that it can do the same or almost the same at a fraction of the priceas its Western competitors.
From a utility point of view, the key feature of the new version of DeepSeek is support for a context window with a length of 1 million tokens. To put it very simply, the context window is responsible for how much the language model can temporarily “remember” when generating responses, which directly translates into its effectiveness in performing complex tasks that operate on large amounts of data.
The number 1 million is important here because this is the size of the context window of the latest flagship Anthropic models (Opus 4.6 and the brand new 4.7), and in the case of OpenAI models you have to pay extra for a window of this size.
Of course, the size of the context window alone is not an indicator of the final quality of the language model, just as engine power says little about how a car will handle on a winding race track. However, the test results published by DeepSeek look promising:
The first analyzes published by external analysts also look equally promising. For example, a team of engineers from FundAI, one of the leading publications writing about the AI market, simply wrote:
— “At the forefront of pioneering models is currently underway three-way competition between Anthropic, DeepSeek and OpenAI, each of which excels in a different field.”
Their tests show that although overall Anthropic still has the most powerful model, its advantage over the rest of the group is decreasing, and for some specific applications, DeepSeek is simply better.
DeepSeek is fighting again on price
This result is interesting in itself, but it becomes even more interesting when you look at the price lists of the creators of flagship language models:
| Output cost (1M tokens) | Context window | |
|---|---|---|
| DeepSeek V4-Flash |
$0.28 |
1 million |
| DeepSeek V4-Pro |
$3.48 |
1 million |
| Claude Opus 4.7 |
$25.00 |
1 million |
| GPT-5.5 |
$30.00 |
1 million |
This combination of potentially great possibilities with a low price, compared to OpenAI and Anthropic, is mind-boggling. Especially now when users of flagship models of Western giants regularly complain about rising costs. OpenAI increases the cost of token generation with each new GPT-5 subversion – when moving from version 5.4 to 5.5, it increased by as much as 1/3.
However, Anthropic works in a different way – on the one hand, it limits the limits of subscription users, de facto forcing many of them to pay for used tokens, and on the other hand, in the latest version of the Opus 4.7 model, the so-called tokenization (i.e. the way the model internally divides work into parts), which means that the same commands now consume much more computing power, which also means the client's money.
We need more details to give DeepSeek a final rating
However, before we start talking about DeepSeek being the “killer” of OpenAI and Anthropic, there are a few details to consider.
Firstly, the quality of the new DeepSeek must be carefully verified and tested in practice. Especially on complex, multi-step tasks that take full advantage of large context windows. When AI agents are given multi-step work to complete, the chances of making a mistake at each step quickly accumulate. This means that a small difference in effectiveness at the beginning may translate into a huge difference in effectiveness, for example at step 20, and make it so that out of two theoretically similar models, one can be implemented and relied on, while the other is useless. This efficiency in performing complex tasks is currently one of the main reasons for Anthropic's success among corporate clients, and without it DeepSeek will not be an “Anthropic killer”.
Secondly, models differ in the efficiency of token use. The same query may require a different number of tokens depending on the model, so ultimately the price difference between DeepSeek and competing models may be different (in favor or against the Chinese model).
Thirdly: In many applications, apart from the price and quality of generated responses, what is important is the so-called interactivityi.e. the time needed to obtain the first token and complete the task. It is particularly important, for example, in programming applications. The historically low price of the official version of DeepSeek was related to the low interactivity of the service, and many customers are now willing to pay many times more to improve the speed of artificial intelligence. These customers will not abandon Anthropic or OpenAI for DeepSeek if it is a little better and much cheaper, but at the same time much slower. However, a lot depends on the specific service and the infrastructure it runs on, and you must remember that DeepSeek is an open model and anyone can run it in their server room.
DeepSeek remains one of China's most popular language models, although it now faces much more competition than it did a year ago.
|
mundissima / Shutterstock
For now, we need more time and information because the situation is not black and white. However, the very fact that it makes sense to compare the new, relatively cheap Chinese language model with the American ones is food for thought. If subsequent comparisons are favorable for DeepSeek, it may pose a threat to the business model of OpenAI and Anthropicbecause these companies are currently able to set higher prices for their services, among other things, because they are the only providers of models that can perform certain complex tasks.
Read also: Trillion dollar questions. The answers to these questions will determine the future of technology
Even if DeepSeek does not defeat the OpenAI and Anthropic models in a direct fight, and the position of these companies among business customers remains unthreatened, the new version of the flagship Chinese model will reinforce a certain trend that has been visible for several months. The trend of popularizing Far Eastern artificial intelligence among less demanding users, for whom it is increasingly “good enough”.
There is one more important aspect of this situation. For now, we don't know yet what hardware DeepSeek was trained on and what it is running on. If these are Huawei accelerators, the issue of the effectiveness of American sanctions will certainly come back.
