Today, DeepSeek did something quite outrageous in the AI circle:
Officially released V4.1 Flash, then announced — V4 Pro will be discontinued in four days (September 14 at 12:00), with all requests automatically routed to the new Flash, billed at Flash prices.
Notice this hierarchy: Flash is the lightweight version, Pro is the flagship version. The younger brother is launched, and the elder brother is retired.
First, let's review how magical this month has been
On August 13, DeepSeek released the official version of V4 Pro, along with a price increase notice: the API price raised up to 11 times, introducing peak and off-peak pricing, doubling the price during peak hours. The developer community erupted, with the words "no value for money" echoed everywhere.
On August 17, the new prices took effect, bringing a wave of complaints.
On September 9, DeepSeek quietly announced a price reduction for the Flash series: the cache hit price was cut by 60%, with many prices returning to pre-increase levels.
On September 10, the truth was revealed: it wasn't a conscience-driven price reduction, but rather a new model that genuinely reduced costs. The V4.1 Flash was released, significantly outpacing V4 Pro in performance, speed, cost, and total time — according to the official statement.
In other words, the official version of V4 Pro was only available on the market for 28 days before being sent to the museum by its own "smaller-sized" model.
How did it manage to be both cheaper and stronger?
Let me mention two technical details, which are easy to understand without technical expertise.
First, the new model changed its brain structure. Previously, the model used the same set of parameters for reading and writing; V4.1 Flash changed to "asymmetric": it activates 8 billion parameters when reading inputs and 16 billion when writing outputs.
Why this change? Because the work mode of AI has changed. An agent running tasks needs to read a lot of code, documents, and tools to return results, while the content it actually writes is relatively little — a typical "read a lot, write less." Since the reading portion doesn't need to be so smart, it uses a smaller brain for reading and a bigger brain for writing, naturally bringing down the bill.
Second, it has compressed the memory footprint of "short-term memory" to the extreme. The KV Cache (which can be understood as the model's scratch paper for reading long texts) has been reduced to 890 bytes per token, with memory requirements dropping to a quarter of the previous generation, and shrinking 437 times compared to DeepSeek's initial model. Agents repeatedly read the context, and this scratch paper was previously a major cost; now it is basically free.
As shown in the picture, DeepSeek has made continuous progress in reducing context storage. The KV Cache has shrunk 437 times compared to the initial model.
Total parameters: 552 billion, MIT license open-source, 1 million tokens for context, and it natively includes visual understanding — no need to adjust models separately for seeing images, reading screenshots, and analyzing charts.
Let's calculate the price
During off-peak hours, each million tokens: cache hit input costs 2 cents, miss costs $1, and output costs $4. Prices double during peak hours.
Compared to the old Flash before the price increase in August: the input price has returned to the previous level, while the output is still twice as high as before. But compared to the price of V4 Pro, someone on Hacker News calculated the dollar figures—output tokens: new Flash costs $0.6 during off-peak, while Pro costs $1.98, more than three times the difference.
So after September 14, a very magical scene will appear: you send requests to the deepseek-v4-pro interface, but in the background, a stronger model runs for you, charging you less.
Downgrade? No, this is called "downward upgrading."
What is more worth pondering is the latter half of the sentence
The official said that V4.1 Flash is "the smallest model in the series of completely new model structures."
The smallest one has already taken down the previous flagship.
What about the larger ones? The report states very clearly: the new structure is designed "to be scalable to larger parameter models." V4.1 Pro is brewing in the shadows. DeepSeek's approach equivalently releases a smaller version first to clear the scene, directly shattering the price anchor of "flagship capabilities," waiting for the larger version to emerge, where the pricing and perceived value will all be dictated by it.
The timing is also very subtle. Just yesterday, Reuters reported that DeepSeek engaged CITIC Securities to assist with its IPO on the Science and Technology Innovation Board, with a valuation of 500 billion; today, it uses a "greater quantity and lower price" event to maximize goodwill among global developers — the related posts on Hacker News surged to nearly 400 points, with comments filled with "ridiculously cheap."
The capital market requires imagination to tell stories, while the developer ecosystem needs genuine affordable prices. DeepSeek has fed both sides within two days.
Looking across the ocean: on the same day, OpenAI's GPT-6 Astra announced it might pause new Pro subscription registrations due to overwhelming demand — one is too busy to keep the doors open, while the other is wide open and reducing prices.
The AI competition is headed to the fall of 2026, and the direction is already very clear: some are selling luxury goods, while others are selling tap water.
For domestic developers, the advice is practical: in the next few days, try to shift Agent tasks to off-peak hours (evenings and weekends) to save half; after September 14, if you are still using the v4-pro interface, you don’t need to change the code, but remember to check the billing in the background — there will be surprises.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
