DeepSeek V4-Pro Announces Permanent Price Drop: API Output Less Than NT$30 per Million Tokens

May 22, 2026, DeepSeek announced that its V4-Pro model API will be permanently discounted at 6 RMB. As AI agent software is being widely deployed worldwide, with explosive demand for computing power, most AI companies are choosing to raise prices, but DeepSeek is going against the trend.
(Background: DeepSeek V4 release—can its programming capabilities beat GPT and Claude? Costs once again top the charts)
(Additional context: Why does China's extreme cost-performance ratio in AI make Silicon Valley feel anxious?)

DeepSeek announced this week that it is making its limited-time discount permanent. After the promotion ends on May 31, the price for every million tokens output will remain at 6 RMB (about 28 NTD), forever.

Note: The API for DeepSeek V4-Pro was previously in a promotional period, with a discount of 25% off the original price. According to market expectations, the promotional activity was set to end in June 2026, and prices would "return to the original rate."

The input side remains low as well. According to this pricing announcement, the input pricing for V4-Pro also stays at the discounted level, placing the overall API rate among the lowest tiers in mainstream commercial models. For ordinary individual users, DeepSeek’s website and app are currently free. The direct beneficiaries of this price cut are developers and enterprise users, especially those with high call volumes and sensitivity to token costs in commercial applications.

Decreasing compute costs, counter-cyclical strategy

Why does DeepSeek choose to permanently lower prices while others raise them? There are three levels of answers to this question.

First layer: Structural decline in computing costs. The DeepSeek V4 series has been confirmed to run entirely on Huawei Ascend 950 PR chips, breaking away from direct reliance on NVIDIA GPUs.

In the context of China’s self-built computing ecosystem, DeepSeek’s inference cost curve is asymmetric compared to Western AI companies. When OpenAI and Anthropic need to bid high for H100/H200 hardware in the market, DeepSeek’s hardware procurement path is different, and its cost structure allows it to adopt more aggressive pricing.

Second layer: Market share strategy. API price reductions are the most direct way to attract developer ecosystems. In April, DeepSeek V4-Pro demonstrated the ability to compete head-to-head with GPT-4o and Claude Sonnet series in coding capability tests.

As the gap in model capabilities narrows, pricing differences directly determine which platform becomes the default choice for developers. Being 36 times cheaper per million tokens means that for commercial applications with tens of millions to hundreds of millions of token calls, the annual API cost drops from millions of dollars to tens of thousands, marking a critical decision point.

Third layer: Deeper signaling. Turning a temporary promotion into a standard price is a market commitment gesture. DeepSeek is telling developers: you can incorporate this pricing into your business plans without it suddenly going back up. This is especially important for enterprise clients who need long-term planning.

Combined, these three layers present a clear strategic image: DeepSeek is using pricing tools to build developer dependency, aiming to make domestically produced foundational models the underlying inference engine for global commercial AI applications.

How big is the gap: 36 times vs 200 times, numbers speak

OpenAI GPT-5.5 standard version costs $30 per million tokens output, roughly 216 RMB. vs DeepSeek V4-Pro’s 6 RMB, a difference of over 36 times.

GPT-5.5 Pro costs $180, roughly 1,296 RMB. vs the same 6 RMB, over 200 times difference.

If a startup’s monthly AI inference cost on OpenAI is $100k, switching to DeepSeek V4-Pro’s theoretical rate could reduce it to below $2,800 (assuming model capabilities are interchangeable in their respective use cases).

Although GPT-5.5 and Claude Opus still have their respective moat in complex reasoning, multilingual understanding, and long-text processing, for many enterprise AI scenarios—customer service, document processing, coding assistance, data summarization—DeepSeek V4-Pro’s performance is already sufficiently strong.

In 2026, when computing costs are generally rising, a permanent price cut is an aggressive political statement: DeepSeek aims to establish its position in the global AI infrastructure layer, and pricing is the fastest way to get in.