Summary of Key Points
After the official release of the domestic open-source AI model DeepSeek V4 Flash, there has been a surge in usage on two major overseas model calling platforms (OpenCode and OpenRouter). It processes 8 trillion Tokens per day and 7.22 trillion Tokens per week, with a 30% increase in users. Its performance is comparable to that of OpenAI's lightweight model Luna (an intelligence score of 50 vs 51), but with a 98% cache discount, its cost is 60% lower than that of Luna even after a 80% price reduction, making it very popular among overseas developers. This phenomenon indicates that the competition between open-source and closed-source models has entered a new phase dominated by cost-effectiveness, as domestic open-source models are challenging the overseas market and putting significant pressure on OpenAI.
I. DeepSeek V4 Flash's Overseas Success: A Double Surge in Usage and Users
The data from the release of DeepSeek V4 Flash is impressive:
- OpenCode Platform: Processes 8 trillion Tokens per day (5 trillion for free trials, 3 trillion for paid usage), with a 30% increase in both usage and new subscriptions.
- OpenRouter Platform: Handles 7.22 trillion Tokens per week, instantly ranking first on the weekly charts.
Although these two platforms serve different purposes (OpenRouter is a general model aggregation gateway, while OpenCode focuses on code development), the significant increase in usage shows that DeepSeek V4 Flash is not only well-received in coding scenarios but also meets a wide range of AI development needs, truly appealing to overseas developers.
II. Head-to-Head Competition with OpenAI Luna: Similar Performance, but a Huge Price Difference
To compete, OpenAI reduced the price of Luna by 80% (from $6 per million Tokens to $1.2, and from $1 to $0.2). However, developers were still not convinced, with many commenting that they wanted the price of DeepSeek.
The reason is simple: the performance difference is minimal, while the cost is significantly lower.
- Intelligence Score: DeepSeek V4 Flash scores 50, just one point behind Luna’s 51, indicating nearly equivalent processing power.
- Cost Comparison: Even after the price cut, DeepSeek's cost per task is still 60% lower than Luna’s. For example, if the input hits the cache, DeepSeek costs only $0.2 per million Tokens, compared to Luna’s $0.2 per million Tokens.
III. The Power of Cost-Effectiveness: A 98% Cache Discount Turns Repeated Calculations into a Cost-Saving Advantage
DeepSeek’s ability to offer lower prices is due to its high cache hit rate discount.
What is a cache? Simply put, if a developer has asked a similar question before (e.g., with repeated code snippets or common issues), the server already has the calculated results available, and there’s no need to recalculate. In this case, only 2% of the Tokens are charged for this part of the process (typically 10% in the industry).
For example, if you use DeepSeek to process 1 million Tokens, and 980,000 of them hit the cache, you would only need to pay $0.396 ($980,000 * $0.2 + $20,000 * $1). Using Luna, even if all hits the cache, it would cost $0.2 (about $1.4), which is nearly three times more.
IV. The New Phase of Open-Source vs Closed-Source Competition: Domestic Models Breaking the Monopoly with Cost-Effectiveness
Previously, OpenAI dominated the overseas AI model market, but now domestic open-source models like DeepSeek are changing the landscape:
- Fast Iteration: DeepSeek V4 Flash’s performance is already on par with OpenAI’s mainstream models.
- Clear Price Advantage: With innovative measures such as cache discounts, DeepSeek reduces costs to less than 40% of those of its competitors.
- Shift in User Preferences: Developers no longer solely rely on OpenAI; they value a combination of good performance and low prices.
OpenAI’s 80% price cut has not stopped the rise of DeepSeek, demonstrating that domestic open-source models have turned cost-effectiveness into a powerful competitive tool, forcing the overseas market to reevaluate China’s AI capabilities.
V. The Significance Behind This: Open-Source Models are Reshaping the AI Industry
The success of DeepSeek is not accidental; it reflects a trend in the AI industry where open-source models are challenging the dominance of closed-source models.
Closed-source models (like OpenAI) rely on “black box” technology and brand premium to generate profits, while open-source models (like DeepSeek) offer similar performance at lower costs through open technology and cost optimization. This not only benefits developers but also promotes technological adoption across the industry. After all, more people needing access to AI will drive more innovative applications.
In the future, AI competition may no longer be about who has the best technology, but about who can deliver that technology efficiently and at a lower price. The successful overseas launch of domestic open-source models is a clear testament to this trend.