On May 31, 2026, the promotional discount on DeepSeek V4 Pro's API pricing expires. But DeepSeek has announced it will not restore the original rates. The 75% discount is hardening into the permanent list price. A service that was paying 1 million won (about $700) a month in API charges will now be billed 250,000 won (about $180) for the same usage. That number isn't hypothetical — it shows up on June's first invoice.
On the developer forum Hacker News, the announcement drew 244 points and 145 comments. What's striking is that many commenters weren't simply cheering the discount — they were already starting to redesign their service architectures around this price. The baseline for AI API costs has genuinely dropped another notch.
The Promotion Ended — and Prices Didn't Go Back Up
The DeepSeek V4 Pro API discount was announced from the start as a temporary measure, with a clearly stated expiration of May 31, 2026, at 15:59 UTC. Most users expected prices to snap back after that point.
In mid-May, DeepSeek announced a different decision through its official documentation and social channels: even after the promotion ended, the quarter-of-original pricing would be officially confirmed as the permanent list price.
Set that number alongside other major models and its position becomes clear. As of May 2026, OpenAI's GPT-4o costs $5 per million input tokens and $15 per million output tokens. Anthropic's Claude Sonnet line runs around $3 for input and $15 for output. Google's Gemini 1.5 Pro offers a free tier up to a certain usage level, with paid pricing beyond it. Compared directly against these, DeepSeek V4 Pro's permanently locked price is dramatically lower.
On performance, DeepSeek V4 Pro posts numbers that compete with the top models from OpenAI and Anthropic on public benchmarks for coding, math, and complex reasoning. The price has dropped at the same moment the benchmarks sit near the top of the leaderboard.
Behind DeepSeek's ability to sustain this price is a difference in architectural design. When DeepSeek R1 was released in early 2025, claims circulated that its training costs were a small fraction — dozens of times lower — than OpenAI's. The exact figures were never independently verified. But the strategy of building price competitiveness on a low inference-cost structure has been consistent throughout. With Google, Amazon, and Meta also steadily cutting prices on their own AI services, DeepSeek's decision can be read as a move to lock in the floor price first in a competitive race.
Draw in API users with low prices, and once those users start designing their services around your platform, they find it hard to leave even if prices rise later. That's why this price lock is something more than an extended promotion.
The Concerns That Come With a Low Price
There are reasons this change is hard to welcome without reservation.
DeepSeek is headquartered in China. Transparency about which servers data sent through the API passes through, and how it is processed, remains insufficient. Some public institutions in Europe and some companies in the US have already restricted DeepSeek internally or issued guidance discouraging its use. How this API can be used under Korea's Personal Information Protection Act — the country's core privacy law — and under the data-handling rules governing finance and healthcare is a question that requires separate legal review.
The pattern of undercutting on price to capture a market and then changing the terms later has repeated throughout the cloud industry. AWS attracted startups in its early days with low prices and a free tier, then adjusted pricing in stages once market dependence had deepened. When a service built deeply on one cloud faces a price hike, switching costs make it hard to move. There is no guarantee, at present, that DeepSeek won't walk the same path.
Service availability also needs verification. During DeepSeek's period of explosive growth in 2025, there were episodes when the API became unstable under surging traffic. Adopting it in a production environment without directly testing real-world latency and availability is a risk of its own.
These concerns don't carry equal weight in every situation. The size of the risk depends on the nature of your service's data and its regulatory environment. Skipping the evaluation just because concerns exist, and adopting without verification just because the price is low, each generate costs in their own way.
Now Is the Time to Pull Out Your Invoices
The people for whom this price change actually matters are those already running AI APIs in real work. If you're a solo business owner in Korea, a one-person product manager, or a middle manager driving AI adoption, there are things to check right now.
Start by figuring out which APIs you're using, and at what volume. Surprisingly often, these costs are lumped together as \"the ChatGPT subscription\" or \"AI tool expenses.\" If you don't know, line by line, which model serves which purpose, exploring alternatives is nearly impossible. Pulling out the last three months of invoices and breaking them down by line item is the starting point. For example, if you're handling customer email summaries and social media draft writing with the same model, you can first calculate what changes when you move the repetitive, less complex work — like drafting — to a cheaper model.
Not every task needs the highest-performing model. Teams have reported cutting costs by 30–60% by routing repetitive work — simple classification, keyword extraction, draft summaries — to inexpensive models, and reserving premium models for complex reasoning or final customer-facing output. With costs down, now is also the moment to redesign that structure from scratch.
To decide whether to adopt DeepSeek, first sort out the nature of the data you process. Internal workflow automation, content generation from public information, and processing that involves 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 heavily regulated sectors like finance or healthcare, legal review must come first.
When smartphones first went mainstream, the teams that explored the market and designed services during the app store's first two or three years held the advantage in the competition that followed. The same thing repeated when cloud services became an option for startups. The people who pulled ahead in those moments had something in common: accumulated, concrete experience experimenting with how a new environment connected to the way they worked. Simply knowing how to use a technology, and judging which technology to connect, when, and how, are entirely different capabilities.
The fact that AI API costs have hardened at a quarter of what they were is a signal that the window for making that judgment has opened again. If you're working right now without a grasp of your own tool-cost structure, that blind spot is quietly draining unnecessary spending out of your bill every month.
