In late April, DeepSeek slashed V4-Pro prices by 75% as a limited-time promotion, and everyone expected prices to rebound at the end of May. On May 23, the company posted in its API console that after the promotion ends, the listed price will remain at the current level—not temporary, but permanent.
V4-Pro's per-million-token pricing has been locked at $0.435 for input and $0.87 for output, down from $1.74 and $3.48 respectively. In RMB, that ranges from 0.025 yuan to 6 yuan per million tokens, cutting the entire price line to a quarter of the original.
Limited-Time Becomes Permanent: A Different Game
In China's AI circle, API pricing promotions have followed the same pattern for years: launch a new version, offer a limited-time discount to attract users, revert to original pricing when the period ends, then discount the next version. Short-term low prices correspond to capacity ramp-up and user acquisition, while corporate financial models still run on original prices—because no one dares to budget for a price that might rise next month.
DeepSeek has reversed this. It uses the promotional price as an anchor, and the long-term price will follow that line. This means the company has calculated that at $0.87 per million output tokens, it can cover hardware depreciation, electricity, and R&D amortization. It is publicly telling all developers that this is the long-term price, safe to include in procurement contracts.
For buyers, this is what truly matters. Temporary promotional prices are too risky for important projects; permanent prices can be used.
$0.87/M Output: Western Rivals Barely Exist at This Level
Comparing frontier models' output token prices:
- Claude Opus 4.7: $75 per million tokens
- GPT-5.5: $10
- Gemini 3.5 Pro: ~$10
- DeepSeek V4-Pro (permanent price): $0.87
That's nearly 90 times cheaper than Claude Opus and over 11 times cheaper than GPT-5.5. The cache hit price for input is even more aggressive—$0.003625 per million tokens, nearly zero. For scenarios like RAG heavy retrieval or multi-turn agent conversations that repeatedly feed context, the more cache used, the more savings.
In benchmarks, V4-Pro still trails Claude Opus 4.7 and GPT-5.5, but the gap has narrowed from a full generation to slightly weaker within the same generation. For many enterprise use cases—customer service, data processing, batch generation, internal knowledge Q&A—the performance gap is small enough that the cost savings can fund hiring two more engineers.
Huawei Is the Real Backbone Behind This Price
DeepSeek's ability to cut prices so aggressively is not unilateral concession. The V4 series was designed from the start to adapt to Huawei's Ascend 950 line. The Ascend 950PR just entered mass production last month, and Huawei expects its AI chip revenue to double from $7.5 billion last year to $12 billion this year—the deeper V4 and similar large-model customers use it, the more Ascend shipments grow.
Bypassing NVIDIA GPUs means DeepSeek's hardware costs are not affected by U.S. export controls and are no longer tied to NVIDIA's generational upgrade pace. Jensen Huang told CNBC's Sara Eisen last week: "The demand in China is quite large. Huawei is very, very strong... We've really largely conceded that market to them." The NVIDIA CEO admitting that China's market is largely conceded to Huawei would have been unthinkable two years ago.
DeepSeek runs V4-Pro on its self-developed Ascend stack, keeping per-token marginal costs low enough to sustain this pricing. Running the same model on NVIDIA H100/H200 would be unprofitable at this price.
Who Will Feel the Pressure Next?
After the price list was published, domestic rivals must respond first. Qwen3.6 is just beginning to roll out commercial versions, and Hunyuan 3.0 was released only in April—both were priced at industry averages. Now DeepSeek has set the anchor at one-quarter of the original. If domestic competitors don't follow, developers will migrate; if they do, they will lose money.
The international market is more nuanced. Claude API and GPT API sell on capability, compliance, and enterprise support, not just price. But token-heavy users—educational products, customer service AI, batch data processing—are already calculating: moving non-core scenarios from Claude/GPT to V4-Pro saves enough to hire an engineer dedicated to adaptation.
The second wave of pressure will hit non-Chinese vendors using NVIDIA. Not because their models are bad, but because their cost curves are on a completely different scale from the "Ascend + cheap domestic electricity" combination.
April's 75% discount was a temporary promotion, and the market could treat it as such. This permanent move publicly sets that pressure line—anyone whose cost structure can't reach this level will have to bow out of the price comparison game with DeepSeek.
Sources: CocoLoop, China's DeepSeek Permanently Cuts V4-Pro AI Costs by 75% (Times of AI); DeepSeek cuts V4-Pro API pricing permanently as developers weigh cost (Business News Today); Models & Pricing (DeepSeek API Docs); Nvidia says it has 'largely conceded' China's AI chip market to Huawei (CNBC)