DeepSeek V4: China rolls out its open-source weapon against OpenAI and Anthropic
On 24 April 2026, DeepSeek released two extremely powerful models under MIT, without Nvidia, and 7 times cheaper than Claude. Decoding the earthquake.
Alyriaz··6 min read
Advertisement
What is DeepSeek V4?
On 24 April 2026, the Chinese lab DeepSeek released two new AI models: V4-Pro (the big brother, ultra-powerful) and V4-Flash (the little brother, faster and cheaper). Both are available immediately on Hugging Face under the MIT licence, which means any developer in the world can download them for free, modify them and use them commercially, without asking anyone for permission.
To grasp the scale of the event: DeepSeek released V4 on the same day OpenAI launched GPT-5.5. That is no coincidence. It is a frontal attack.
Why is everyone talking about it?
Three reasons, all three of them uncomfortable for Silicon Valley.
First, the price-performance ratio. V4-Pro costs $3.48 per million output tokens, against $25 for Claude Opus 4.7 and $30 for GPT-5.5. That is a price gap of more than 7 times for performance that stays within a few points on most benchmarks.
Second, fully open source. You can download the model weights, run them on your own servers, and share nothing with anyone. That is radically different from Claude or GPT-5.5, which remain closed and accessible only through an API controlled by their creators.
Finally, the technological break with hardware. For the first time, a frontier-class model has been trained and optimised on Huawei Ascend 950PR chips, not on Nvidia. DeepSeek even turned down early access to Nvidia in order to give Huawei that window exclusively. Jensen Huang, Nvidia's CEO, described the situation as a "disaster" for the United States in an interview with the Chinese press.
The technical innovation that changes everything 🧠
V4 uses a new architecture called CSA + HCA (Compressed Sparse Attention and Heavily Compressed Attention). Put simply: it is as if you had to retain an entire 1,000-page book, and instead of memorising it word for word, you kept whole chapters in summary and the important sentences in detail. The result: V4 handles contexts of 1 million tokens using only 27% of the computing power and 10% of the memory of its predecessor V3.2.
How does V4-Pro compare with Claude Opus 4.7 and GPT-5.5?
The three flagship models came out in the same week of April 2026: Claude Opus 4.7 on the 16th, GPT-5.5 and DeepSeek V4-Pro on the 23rd-24th. Here is the verdict on the benchmarks directly comparable between the three:
Benchmark
DeepSeek V4-Pro
Claude Opus 4.7
GPT-5.5
Terminal-Bench 2.0 (command-line tasks)
67.9%
69.4%
82.7%
SWE-Bench Pro (fixing real GitHub bugs)
55.4%
64.3%
58.6%
BrowseComp (autonomous web browsing)
83.4%
79.3%
84.4%
GPQA Diamond (scientific reasoning)
90.1%
94.2%
93.6%
HLE without tools (multi-domain expert questions)
37.7%
46.9%
41.4%
On the directly comparable benchmarks, DeepSeek V4-Pro does not beat either of the two closed models, but remains very close on autonomous web browsing (BrowseComp) and the terminal. On complex reasoning (HLE, GPQA Diamond), Claude Opus 4.7 retains a solid lead. DeepSeek's real argument is not absolute performance, it is value for money.
And against the other open-source models?
DeepSeek V4 is not just fighting the American giants. China's open-source market has become hyper-competitive in 2026, with two direct competitors: Kimi K2.6 from Moonshot AI and Qwen 3 Max from Alibaba.
Open-source model
SWE-Bench Pro
Context
Licence
DeepSeek V4-Pro
55.4%
1M tokens
MIT (free)
Kimi K2.6 (Moonshot)
58.6%
200K tokens
Modified
Qwen 3 Max (Alibaba)
~52%
1M tokens
Apache 2.0
Kimi K2.6 remains slightly ahead of V4-Pro at fixing real-world bugs, but V4 regains the advantage on competitive programming and long context. For the vast majority of enterprise use cases, V4-Flash (the "small" version) is more than enough and costs even less.
China's rise is no longer a threat, it is a reality
A year and a half ago, in January 2025, DeepSeek R1 had wiped nearly 600 billion dollars off Nvidia's market value in a single day. At the time, some analysts spoke of a "Sputnik moment" for Chinese AI, a reference to the Soviet satellite that caught America off guard in 1957.
With V4, it is no longer an isolated moment. It is a trajectory. China is methodically building a complete, independent chain: Chinese open-source models, Chinese chips (Huawei), Chinese inference software. Alibaba, ByteDance and Tencent have already placed orders for the next generation of Huawei chips to host V4 on their cloud services.
>>
For companies and researchers operating in mainland China, this technological autonomy probably matters more than any benchmark score. For the rest of the world, it is the signal that a credible alternative to the Nvidia-OpenAI-Anthropic ecosystem now exists.
The geopolitical context 🌍
The release of V4 comes as the White House accuses Beijing of "industrial theft" of AI technologies, and as a Trump-Xi meeting looms. Demonstrating Chinese technological parity at the exact moment Washington is trying to tighten controls on chip exports is probably not a coincidence of timing.
How much does it really cost?
The DeepSeek V4 API pricing is what makes everything else possible:
Model
Input (per M tokens)
Output (per M tokens)
DeepSeek V4-Pro
$0.55
$3.48
DeepSeek V4-Flash
$0.28
$1.12
Claude Opus 4.7
$5
$25
GPT-5.5
$5
$30
In concrete terms: for an application generating 10 million tokens a month (the equivalent of an active professional chatbot), V4-Pro will cost you around $40 a month. The same usage with Claude Opus 4.7 would set you back $250. With GPT-5.5, $300. For an independent developer or a small team, that is the difference between a project being viable and being abandoned.
Which model for which use case?
No need to overcomplicate things, here are the three typical cases:
>>
For absolute quality (critical production): Claude Opus 4.7 remains the best choice for code reliability and complex reasoning, despite its price.
For automated workflows (agents): GPT-5.5 keeps the edge on chaining long tasks and tool use.
For everything else (90% of real-world uses): DeepSeek V4-Flash. At under $1 per million input tokens, with a 1-million-token context window, it is simply the best value for money on the market in April 2026.
Has open source just won the war?
Not quite. Claude Opus 4.7 retains concrete advantages, notably on complex reasoning and tasks that demand absolute reliability. GPT-5.5 remains unbeatable on agentic autonomy. But for the first time, an open-source model is seriously competing with the closed giants on their own benchmarks, at 7 times less cost.
The stakes are no longer about raw performance. They are about what companies are prepared to accept in terms of technological dependency. Running DeepSeek V4 on your own servers, without going through OpenAI or Anthropic, without depending on an Nvidia chip, without sending your data to a foreign provider: that is now a realistic option. How will this new landscape redraw the balance of power between America and China on AI in the coming months? That is precisely what we will explore in a forthcoming article.