DeepSeek V4 Launches with Flash and Pro Versions, Pricing Unveiled
A major update has arrived in the large model space. Leading Chinese AI company DeepSeek today launched its new flagship model DeepSeek V4. The key highlight is its differentiated strategy, which addresses diverse needs—from high-frequency lightweight tasks to complex reasoning—through two versions: Flash and Pro. With aggressive pricing, it once again sets a new benchmark for AI commercial costs.
Model Matrix: Differentiated Positioning of Flash and Pro
DeepSeek V4 integrates and upgrades the original deepseek-chat and deepseek-reasoner models, officially splitting them into two versions:
DeepSeek-V4-Flash: Focused on extreme cost-efficiency and high throughput, ideal for fast-response general conversations and basic text tasks.
DeepSeek-V4-Pro: Optimized for complex logic, deep reasoning, and high-performance computing, offering stronger reasoning and processing capabilities.
Both models support thinking mode (except in specific scenarios), JSON output, tool calls, dialogue prefix continuation (Beta), and a context window of up to 1M tokens with a maximum output of 384K tokens, providing a solid foundation for complex engineering implementations.

Pricing System: Transparent, Tiered Billing
DeepSeek’s pricing is clear and significantly reduces the marginal cost for long-term enterprise API calls through a caching mechanism. Here are the billing rates per million tokens (in RMB):
ModelInput (Cache Hit)Input (Cache Miss)OutputDeepSeek-V4-Flash0.2 RMB1 RMB2 RMBDeepSeek-V4-Pro1 RMB12 RMB24 RMBNote: Fees are deducted first from the free balance, then from the paid balance.
Industry Analysis: Why This Price Is a Milestone
The pricing logic shows DeepSeek is encouraging developers to reduce compute waste by optimizing cache usage, enabling fine-grained cost control through the significant price gap between cached and uncached inputs.
For developers, the Flash version at 1 RMB per million tokens (cache miss) dramatically lowers the barrier to accessing top-tier MoE model architectures. Meanwhile, the Pro version’s pricing for complex reasoning tasks provides a domestic solution for building knowledge-base Q&A systems, autonomous agents, and other advanced applications, balancing performance and cost.
Important Compatibility Note
The official reminder: the original model names deepseek-chat and deepseek-reasoner will be formally deprecated. To ensure a smooth transition, developers can now call deepseek-v4-flash (corresponding to non-thinking mode) and deepseek-v4-pro (corresponding to thinking mode).
For detailed API information and migration guides, developers can visit the official DeepSeek API documentation for more details.
Related article
Apple Clarifies Gemini Rumors: Siri AI to Be Built In-House
Following the 2026 Worldwide Developers Conference (WWDC), Apple addressed media inquiries regarding rumors that its new AI infrastructure was merely a "Gemini shell." The company clarified that Apple Foundation Models (AFM) are not a rebranding of G
OpenAI Still Struggles to Control Rogue AI Activity
On Friday, OpenAI launched a dedicated page for “misalignment reports,” revealing an alarming volume of rogue behaviors documented over an extended period. The site currently lists nine incidents, the majority occurring during reinforcement learning
Flock CEO claims mics detect only gunshots despite website claims
Flock CEO Garrett Langley stated that the company’s gunshot-detection microphones capture “only gunshots,” yet Flock’s website presents a much wider scope of the ambient audio it records.Speaking at TechCrunch Disrupt, the surveillance firm’s co-foun
Related Special Topic Recommendations
Comments (0)
0/500
A major update has arrived in the large model space. Leading Chinese AI company DeepSeek today launched its new flagship model DeepSeek V4. The key highlight is its differentiated strategy, which addresses diverse needs—from high-frequency lightweight tasks to complex reasoning—through two versions: Flash and Pro. With aggressive pricing, it once again sets a new benchmark for AI commercial costs.
Model Matrix: Differentiated Positioning of Flash and Pro
DeepSeek V4 integrates and upgrades the original deepseek-chat and deepseek-reasoner models, officially splitting them into two versions:
DeepSeek-V4-Flash: Focused on extreme cost-efficiency and high throughput, ideal for fast-response general conversations and basic text tasks.
DeepSeek-V4-Pro: Optimized for complex logic, deep reasoning, and high-performance computing, offering stronger reasoning and processing capabilities.
Both models support thinking mode (except in specific scenarios), JSON output, tool calls, dialogue prefix continuation (Beta), and a context window of up to 1M tokens with a maximum output of 384K tokens, providing a solid foundation for complex engineering implementations.

Pricing System: Transparent, Tiered Billing
DeepSeek’s pricing is clear and significantly reduces the marginal cost for long-term enterprise API calls through a caching mechanism. Here are the billing rates per million tokens (in RMB):
ModelInput (Cache Hit)Input (Cache Miss)OutputDeepSeek-V4-Flash0.2 RMB1 RMB2 RMBDeepSeek-V4-Pro1 RMB12 RMB24 RMBNote: Fees are deducted first from the free balance, then from the paid balance.
Industry Analysis: Why This Price Is a Milestone
The pricing logic shows DeepSeek is encouraging developers to reduce compute waste by optimizing cache usage, enabling fine-grained cost control through the significant price gap between cached and uncached inputs.
For developers, the Flash version at 1 RMB per million tokens (cache miss) dramatically lowers the barrier to accessing top-tier MoE model architectures. Meanwhile, the Pro version’s pricing for complex reasoning tasks provides a domestic solution for building knowledge-base Q&A systems, autonomous agents, and other advanced applications, balancing performance and cost.
Important Compatibility Note
The official reminder: the original model names deepseek-chat and deepseek-reasoner will be formally deprecated. To ensure a smooth transition, developers can now call deepseek-v4-flash (corresponding to non-thinking mode) and deepseek-v4-pro (corresponding to thinking mode).
For detailed API information and migration guides, developers can visit
Apple Clarifies Gemini Rumors: Siri AI to Be Built In-House
Following the 2026 Worldwide Developers Conference (WWDC), Apple addressed media inquiries regarding rumors that its new AI infrastructure was merely a "Gemini shell." The company clarified that Apple Foundation Models (AFM) are not a rebranding of G
OpenAI Still Struggles to Control Rogue AI Activity
On Friday, OpenAI launched a dedicated page for “misalignment reports,” revealing an alarming volume of rogue behaviors documented over an extended period. The site currently lists nine incidents, the majority occurring during reinforcement learning
Flock CEO claims mics detect only gunshots despite website claims
Flock CEO Garrett Langley stated that the company’s gunshot-detection microphones capture “only gunshots,” yet Flock’s website presents a much wider scope of the ambient audio it records.Speaking at TechCrunch Disrupt, the surveillance firm’s co-foun





Home






