DeepSeek, the Chinese lab that shook up the AI market in early 2025, has officially launched its V4 model, bringing significant changes to both capabilities and pricing.
A million tokens of context
The most notable technical feature of DeepSeek V4 is its 1 million-token context window across the entire lineup. That puts it on par with Gemini for handling very long documents, large codebases or entire knowledge bases in a single conversation.
The architecture is still Mixture-of-Experts: V4-Pro has 1.6 trillion total parameters but only activates 49 billion per inference, keeping it efficient. V4-Flash, the lighter version, has 284 billion total parameters with 13 billion active.
Peak-hour pricing - a first for the industry
DeepSeek is introducing time-of-day pricing for the first time, a model that resembles electricity billing more than software:
- Peak hours (9:00-12:00 and 14:00-18:00 Beijing time): double price
- Off-peak hours (rest of the day): base price
For European and American users, this could be an advantage: our working hours line up with China’s off-peak hours, which could mean lower costs.
End of the old endpoints
Important for developers: the deepseek-chat and deepseek-reasoner endpoints stop working as of July 24, 2026. Existing integrations must migrate to the new V4 API.
Why it matters
DeepSeek V4 keeps the pressure on OpenAI and Anthropic in the high-performance, low-cost segment. Its open-source models (MIT license) remain the top performers in that category, and this new version improves especially in math reasoning, agent execution and code generation.
For startups and independent developers, DeepSeek remains one of the most affordable options on the market without sacrificing quality.