DeepSeek V4 Returns to Challenge US AI Giants With Aggressive Pricing
DeepSeek achieved worldwide fame in 2025, sending shockwaves through the US-dominated AI industry in particular. At the time, the cost-efficient open-source model V3 was released, which was on par with US models while being significantly cheaper. Since then, a lot has happened, especially in China, and DeepSeek has not been the number one in open-source LLMs for some time; startups such as Z.ai (GLM-5) or Moonshot AI (Kimi K2.6) have overtaken it.
Now, however, DeepSeek is fighting its way back to the front with the release of V4. The two models DeepSeek-V4-Pro and DeepSeek-V4-Flash are available immediately and are being released simultaneously as open-source models. Particularly noteworthy is the combination of high performance and significantly lower costs compared to the leading American competitor models. Currently, V4-Pro ranks just behind GLM-5 in the open-source category on Arena.ai, but only manages 20th place in the overall ranking of all LLMs.
Two Models for Different Requirements
The V4 series consists of two variants that differ in size and performance profile. Both models support a context window of one million tokens and are built on a Mixture-of-Experts architecture. They support both a normal mode and a thinking mode, in which the thinking intensity can be adjusted via the reasoning_effort parameter.
- DeepSeek-V4-Pro has a total of 1.6 trillion parameters, of which 49 billion are active. According to DeepSeek, the model achieves performance comparable to leading closed models, particularly in the areas of mathematics, STEM, and coding competitions.
- DeepSeek-V4-Flash has a total of 284 billion parameters, of which 13 billion are active. It is faster and more cost-effective, showing similar performance to the Pro model on simple tasks, but falling behind on complex tasks.
Performance on Par with the World’s Best
According to its own statements, DeepSeek-V4-Pro outperforms all currently publicly evaluated open-source models in the areas of mathematics, STEM, and coding competitions. In terms of world knowledge, the model falls just short of the leading closed model Gemini-Pro-3.1, but clearly outperforms all other open-source alternatives.
The improved agent capabilities are particularly highlighted. In internal tests, the user experience was rated as better than Claude Sonnet 4.5, and the output quality is said to approach that of Opus 4.6 in normal mode. According to DeepSeek, V4-Pro has already become the internal default model for coding tasks within the company.
Technically, DeepSeek-V4 employs a novel attention mechanism that compresses tokens and is combined with the proprietary DeepSeek Sparse Attention. This enables the long context while significantly reducing computational and memory requirements compared to conventional methods.
The Decisive Advantage: The Prices
What makes DeepSeek V4 particularly interesting is the aggressive pricing. Both models significantly undercut their direct competitors. DeepSeek-V4-Flash is the cheapest model in its class and even beats OpenAI’s GPT-5.4 Nano. DeepSeek-V4-Pro, in turn, is the most affordable of the high-performance frontier models and costs only a fraction compared to Claude Sonnet 4.6 or GPT-5.4.
| Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 |
| GPT-5.4 Nano | $0.20 | $1.25 |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 |
| Gemini 3 Flash Preview | $0.50 | $3.00 |
| GPT-5.4 Mini | $0.75 | $4.50 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
| DeepSeek V4 Pro | $1.74 | $3.48 |
| Gemini 3.1 Pro | $2.00 | $12.00 |
| GPT-5.4 | $2.50 | $15.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
| Claude Opus 4.7 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
The price differences are substantial. Those who use DeepSeek-V4-Pro instead of Claude Sonnet 4.6 pay more than four times less for output. Compared to GPT-5.5 or Claude Opus 4.7, the difference is even more drastic. According to DeepSeek, this is made possible by the high efficiency of the new architecture, which significantly reduces computational and memory requirements, especially for long contexts.
Availability and API Access
Both models are accessible via the official website chat.deepseek.com as well as through the DeepSeek app. The API supports the interfaces of OpenAI ChatCompletions and Anthropic. To access the new models, only the model name needs to be changed to deepseek-v4-pro or deepseek-v4-flash. The previous model names deepseek-chat and deepseek-reasoner will be discontinued in three months, specifically on July 24, 2026.
