Summary in One Sentence:
DeepSeek, the domestic large-scale AI model company that shocked the world with its “rock-bottom” API prices, recently announced a comprehensive price increase of up to 1100% in August. Just 23 days later, it announced a “60% price cut.” However, this was actually just a adjustment of the prices for the widely used Flash model, bringing them back to their initial levels. The prices for the most profitable AI-generated content remained virtually unchanged. Coupled with the news that DeepSeek is preparing for an IPO on the STAR Market, this “price cut” is essentially a strategic move to balance multiple pressures: diverting users to its open-source model versions, managing its “high cost-performance” image, and ensuring both traffic and profits before going public. It’s far from a genuine gesture of giving benefits to users.
---
Explanation in Four Dimensions:
1. Don’t Be Fooled by Marketing: This is a “Fake Discount, Real Fine-Tuning” – Prices Haven’t Returned to Pre-Rise Levels
The so-called “60% price cut” is merely a publicity stunt targeting the least important aspects:
To put it simply:
- Cache hits: These are duplicate pieces of content that the AI has already processed (for example, if you ask the AI to revise the same 100-page contract multiple times, the parts that don’t need to be recalculated are the cheapest to process).
- Cache misses: These are new pieces of content that the AI is processing for the first time and are more expensive to generate.
- Output: The final generated content or code is the most expensive part.
In April this year, the price for 1 million characters of old content was 0.02 yuan, new content was 1 yuan, and AI-generated content was 2 yuan. After the price increase in August, prices doubled during peak times, while old content became 0.05 yuan and new content became 1.5 yuan during off-peak hours, and AI-generated content became 4.5 yuan. The price cut only reduced the prices of old and new content during off-peak hours to their April levels, with a 11% reduction. However, the price of the most valuable output content only dropped from 4.5 yuan to 4 yuan, which is still twice as much as the April price. During peak times, the price of AI-generated content increased to 8 yuan, four times the original amount. The expensive Pro model didn’t see any price reduction at all. It’s like going to a restaurant and the owner says “I’ll give you a 40% discount,” but they only waive the cost of the free appetizer, and the main dish only gets cheaper by 1 yuan, making it actually more expensive than before.
2. The Most Ridiculous Competitor is Themself: Open-Sourcing Has Ruined Their Own Business
Public data shows that DeepSeek’s Flash model is the most frequently used globally, accounting for nearly a quarter of traffic on third-party platforms. So why lower prices when it’s so popular? The main reason is that DeepSeek has made the weights of its Flash model completely open-source, essentially giving away the “secret recipe” for its success. Now, 28 companies around the world can run their own Flash models on servers with almost identical results to DeepSeek’s, but at significantly lower prices.
Developers don’t need to modify any code; they just need to change the API address to use the same model, yet they pay the third-party companies instead of DeepSeek. When prices rose in August, many users switched to cheaper, legally deployed versions of DeepSeek’s models, causing a significant loss of traffic for DeepSeek. If they hadn’t lowered the prices for the widely used Flash model, they would have lost users and thus all the traffic, resulting in no profits.
3. Success and Failure Both Stem from Cost-Performance: DeepSeek’s “Volkswagen Phaeton” Paradox
DeepSeek managed to stand out against overseas giants like GPT and Claude by offering ridiculously low prices. While others priced in dollars, DeepSeek priced in RMB, creating the impression of being “the affordable and high-quality option in the AI industry.” This image was once its greatest strength but has now become a liability. It’s like Volkswagen investing over 10 billion yuan in the Phaeton, which has the same quality as the Mercedes S-Class, yet people associate Volkswagen with being a budget brand and are unlikely to buy a Mercedes at the same price. DeepSeek faces the same issue: without introducing a significantly better new model, a sudden price increase of several times would be unacceptable to users. They lowered the prices for the widely used models to avoid destroying this image and losing existing users.
4. The Price Cut is a Path to IPO: Balancing Traffic and Profit
The day before the price cut, it was revealed that DeepSeek had hired CITIC Securities to prepare for an IPO on the STAR Market. This is the real behind the scenes motivation for the price cut. To go public, AI model companies need to convince investors of two key points: first, they must have a significant market share and be industry leaders (which DeepSeek has already demonstrated with its traffic). If the price increase in August drove users away and reduced its market share, this would undermine this narrative. Second, they must be profitable; otherwise, investors won’t invest.
DeepSeek’s actions are precisely targeted: by lowering the prices for the most sensitive input services, it draws users back to its official platform, maintaining its market share and creating a strong narrative for the IPO. However, it refuses to lower the prices for the most profitable output services, ensuring that each generated word still generates several times more revenue than before the price increase. This strategy not only preserves its traffic but also boosts overall revenue and profits, presenting a perfect image to investors: “high growth, large market share, and improved profitability.” It’s a much smarter approach than simply raising prices and losing users or cutting prices and incurring losses.