DeepSeek V4: Chinese AI Model Undercuts GPT-5.5 by 97 Percent
May 3, 2026
In late April 2026, DeepSeek released its new V4 language model. It runs on Chinese Huawei Ascend chips, supports a one-million-token context window, and charges a fraction of what Western providers ask for input tokens.
What DeepSeek V4 is
On April 24, 2026, DeepSeek released a preview of its V4 language model. It comes in two sizes, V4-Pro and V4-Flash, and is optimized for agentic tasks, knowledge work, and inference. A central novelty is a hybrid attention mechanism that combines two techniques: Compressed Sparse Attention and Heavy Compressed Attention. According to DeepSeek, this allows a one-million-token context window while using 9.5x to 13.7x less memory than V3.2.
A pricing assault on US providers
The biggest headline is price. DeepSeek-V4-Pro starts at around 0.0036 US dollars per million cached input tokens. For comparison, OpenAI's GPT-5.5 charges about 0.50 dollars per million cached input tokens. Multiple outlets describe a pricing advantage of up to 97 percent over OpenAI models.
Huawei instead of NVIDIA
V4 is explicitly tuned for inference on Huawei's Ascend supernode chips. That is more than a technical footnote. It signals that China is steadily assembling a complete domestic AI stack, from chip fabrication through model to application, reducing reliance on US semiconductors such as NVIDIA's.
Why this matters
For businesses worldwide the cost math of AI shifts. If inference for agentic workloads stays this cheap, the entry barrier drops for applications that today are economically unviable: mass classification, continuous monitoring, or long-running research.
Geopolitically V4 signals that the US lead in model quality no longer automatically translates into a price or hardware lead. European businesses will need to look carefully at any AI purchase, including data sovereignty, sanctions, and compliance.
Simple example
A mid-sized manufacturer wants to auto-categorize thousands of customer emails every day. With an expensive US model that adds up to a four-figure monthly bill. With a model as cheap as V4-Pro, the same workload becomes orders of magnitude cheaper on paper. The question shifts from whether AI is worth it to which data is allowed to be processed and where.
What decision-makers should check
Before any deployment comes the privacy and security review. Chinese providers operate under a different legal framework than US providers. Anyone considering V4 or similar models should clarify contracts, hosting location, and data processing agreements before production data enters the picture.
💡 In plain English
Picture a new baker who makes bread almost as good as the priciest one in town, but charges only a few cents. That is what DeepSeek V4 is doing in the AI world. It comes from China, runs on Chinese computers, and costs much less than the well-known US rivals.
Key Takeaways
- →DeepSeek released V4 as a preview on April 24, 2026.
- →V4 costs roughly 97 percent less than OpenAI GPT-5.5 on cached input.
- →The model supports a one-million-token context window.
- →V4 is explicitly tuned for Huawei Ascend chips, not NVIDIA.
- →Privacy and compliance questions need to be settled before deployment.
FAQ
Is DeepSeek V4 open source?
DeepSeek has historically released models with open weights. For V4's exact license check the official model card.
Which chip is V4 built for?
V4 is optimized for Huawei's Ascend supernode inference platform.
How does V4 stack up against US models?
DeepSeek advertises strong performance on agentic tasks and knowledge work. Independent benchmarks against GPT-5.5 vary depending on the test.