On May 31, 2026, the promotional discount on DeepSeek V4 Pro's API officially expired. But DeepSeek announced it would not roll the price back to where it started. In other words, the 75% discount is now locked in as the permanent list price. A service that had been billed the equivalent of one million won a month for the API will now be charged 250,000 won for the same usage. This isn't a hypothetical. It shows up on the first June invoice.
On Hacker News, the developer community, the announcement drew 244 points and 145 comments. What was striking is that a large share of those comments weren't simply celebrating the lower price—they described developers already redesigning their service architecture around it. The baseline cost of AI APIs had, in effect, dropped another notch.
The promotion ended, but the price didn't go back up
DeepSeek V4 Pro's API discount was announced from the outset as a temporary measure. The expiration was spelled out clearly: 15:59 UTC on May 31, 2026. Most users assumed the price would revert after that point.
In mid-May, DeepSeek used its official documentation and social channels to announce a different decision. Even after the promotion ended, it would formally fix the quarter-price rate as its permanent list price.
Lining this number up against other major models reveals where it sits. As of May 2026, OpenAI's GPT-4o runs $5 per million input tokens and $15 per million output tokens. Anthropic's Claude Sonnet line is roughly $3 for input and $15 for output. Google's Gemini 1.5 Pro offers a free tier up to a certain volume and charges beyond it. Set directly against these, DeepSeek V4 Pro's permanent fixed price is markedly lower.
On performance, DeepSeek V4 Pro posts numbers that compete with the top-tier models from OpenAI and Anthropic on public benchmarks for coding, math, and complex reasoning. The price has come down at the same time the performance benchmarks sit near the top of the field.
Behind DeepSeek's ability to hold this price is a difference in architectural design. When DeepSeek R1 launched in early 2025, there were claims that its training cost was a small fraction—on the order of tens of times less—than OpenAI's. The exact figures have not been independently verified. What has stayed consistent, though, is a strategy built on a low inference-cost structure as the foundation for price competitiveness. With Google, Amazon, and Meta all continuously cutting the prices of their own AI services, DeepSeek's latest move can be read as a bid to be the first to set the floor in a competitive market.
Draw in API users with a low price, get them to build their services around the platform, and you create a structure where they can't easily leave even if the price later goes up. That is why this price lock is not a simple extension of a promotion.
The concerns raised in the face of a low price
There are reasons it's hard to greet this change with pure enthusiasm.
DeepSeek is a company headquartered in China. There isn't enough transparency about which servers the data passed through the API travels across, or how it is processed. Some public agencies in Europe and some companies in the United States have already restricted DeepSeek internally or issued guidance to avoid its use. How this API can be applied under Korea's Personal Information Protection Act, and under the data-handling rules of the financial and medical sectors, is a matter that requires separate legal review.
The pattern of capturing a market with low prices and then changing the terms afterward has repeated itself across the cloud industry. AWS drew in startups early on with low-price policies and free tiers, then adjusted prices in stages once dependence on the platform had deepened. When a service that leans heavily on a cloud provider faces a price hike, switching costs make it hard to move. There is currently no guarantee that DeepSeek won't follow the same path.
Service availability also needs to be checked. During the period in 2025 when DeepSeek's service was growing explosively, there were instances where the API became unstable under traffic surges. Adopting it without directly testing real-world latency and availability in a production environment is a separate risk.
These concerns do not apply with equal weight to every situation. The size of the risk shifts with the nature of a service's data and its regulatory environment. Skipping the review simply because concerns exist, and adopting the API without verification simply because the price is low, each generate costs in their own different way.
Now is the time to pull out your invoice
The people for whom this price change actually matters are those already using AI APIs in real, working systems. If you're a solo operator in Korea, a one-person PM, or a middle manager pushing an AI rollout, there are things to check right now.
Start by understanding which APIs you're using and at what volume. More often than you'd think, this gets lumped together as a "ChatGPT subscription" or a "cost of AI tools." Without knowing, line by line, which model you're using for which purpose, it's hard to explore alternatives. Pulling out the last three months of invoices and breaking them down item by item is the starting point. For example, if you're handling both customer-email summaries and social-media draft writing with the same model, you can first calculate how the cost changes when you peel off the repetitive, less complex work—like draft writing—onto a cheaper model.
Not every task needs the highest-performance model. There are reports of teams cutting costs by 30 to 60 percent by handling repetitive work—simple classification, keyword extraction, draft summaries—with cheaper models, and reserving the expensive models only for complex reasoning or the final, customer-facing output. Now that costs have come down is also the moment to redesign this structure.
To judge whether to adopt DeepSeek, you first have to sort out the nature of the data you're handling. Internal workflow automation, content generation based on public information, and processing tasks that contain no personal data carry relatively low risk. If you handle customers' personal information directly, are contractually required to guarantee where data is processed, or operate in a heavily regulated sector like finance or medicine, legal review has to come first.
When the smartphone first went mainstream, the teams that explored the market and designed services in the first two or three years after the app store opened ended up in a favorable position in the competition that followed. The same outcome repeated when cloud services became an option for startups. Throughout, the people who got ahead shared something in common: they had accumulated hands-on experience testing, in concrete terms, how a new environment connected to the way they actually worked. Simply knowing how to use a technology is an entirely different capability from being able to judge which technology to connect, and when, and how.
The fact that the cost of AI APIs has settled at a quarter of where it was is a signal that the window to make that judgment has opened again. If you're working without a grasp of your own tool-cost structure right now, that ignorance is quietly draining unnecessary spending from your invoice every single month.
