The Discount That Wasn’t Temporary
DeepSeek just made its 75% flagship model discount permanent. What started as a promotional price cut is now the new normal — and it’s reshaping the economics of AI for everyone.
Bloomberg first reported the shift on May 23, 2026. DeepSeek’s V4 Pro, which competes with GPT-5.5 and Claude Opus 4 on most benchmarks, will remain priced at roughly one-quarter of its Western rivals’ rates indefinitely. The discount applies to API access, fine-tuning, and batch inference.
On Hacker News, the story hit 42 points and 41 comments in its first hour — a solid showing for a pricing announcement. The comment threads were a mix of grim acceptance from Western AI companies and barely concealed glee from developers. One top comment: “If it’s 75% cheaper and 90% as good, what exactly am I paying OpenAI for?”
🔍 THE BOTTOM LINE
DeepSeek isn’t undercutting on quality anymore — it’s undercutting on price while matching on capability. The AI pricing war has been theoretical for months. It just became structural.
Why This Matters Now
The 75% discount had been running since DeepSeek V4 Pro launched in early 2026. At the time, most analysts assumed it was a land-grab strategy — burn cash to gain market share, then raise prices once you’ve locked in developers. That’s the classic playbook. Uber did it. Amazon did it. Every platform company that wanted to become a monopoly did it.
But DeepSeek isn’t raising prices. It’s cementing them.
This matters for three reasons:
-
It’s not a loss leader if it’s permanent. DeepSeek’s parent company, High-Flyer Capital, has access to cheap compute, custom T-Head chips, and China’s subsidized data centre infrastructure. The discount may actually be profitable — or at least sustainable — at Chinese cost structures. Western labs running on Nvidia H100s and Azure credits can’t match that.
-
It’s a pricing floor, not a ceiling. If DeepSeek V4 Pro costs 75% less than GPT-5.5, every other model’s pricing is now measured against that benchmark. OpenAI, Anthropic, and Google can either cut prices (and margins) or explain to enterprise customers why they’re paying 4× more.
-
It’s open weights. DeepSeek’s models remain downloadable. You can run them locally, fine-tune them, and never pay DeepSeek a cent. The API pricing is just for people who don’t want to manage their own infrastructure. The floor isn’t just cheap API access — it’s free self-hosting.
The Qwen Problem (For DeepSeek’s Competitors)
This announcement lands the same week Alibaba’s Qwen3.7-Max demonstrated a fully autonomous 35-hour kernel optimization achieving 10x speedup — beating DeepSeek V4 Pro’s own score on the same benchmark. China now has two frontier labs producing models that are simultaneously cheap, capable, and getting better fast.
Western AI companies are staring at a two-front problem: DeepSeek on price, Qwen on capability, both with access to Chinese infrastructure economics that make Western margins look indulgent.
The Developer Reaction
The HN thread tells the story:
-
Small developers are switching. A 75% permanent discount on a model that scores within a few points of GPT-5.5 on most benchmarks makes the decision easy. “I moved three production workloads last week,” one developer wrote. “My AWS bill dropped by $4,000/month.”
-
Enterprise teams are more cautious. Data residency, SLA guarantees, and the geopolitical risk of depending on a Chinese provider for core infrastructure are real concerns. But the pricing pressure is forcing OpenAI and Anthropic to justify every dollar of their premium.
-
The “but censorship” argument is getting weaker. DeepSeek’s latest models are noticeably less filtered than earlier versions. They’re not censorship-free — nothing operating under Chinese law can be — but the gap between DeepSeek’s political filtering and OpenAI’s content policies is narrowing, not widening.
What This Means for NZ
New Zealand’s AI adoption has been sluggish partly because of cost. When API access to frontier models costs hundreds or thousands of dollars per month for even moderate use, small NZ businesses — the backbone of the economy — are priced out.
DeepSeek’s permanent pricing makes frontier AI accessible at a price point that works for NZ SMEs. The NZ AI Forum’s refreshed Blueprint for Aotearoa (released this month) talks about accelerating adoption. This is what acceleration looks like when cost barriers drop by 75%.
The catch: data sovereignty. NZ’s Privacy Act and data residency requirements still apply. Businesses using DeepSeek’s API are sending data to Chinese servers. For many use cases (content generation, translation, summarisation), that’s acceptable. For others (customer data, financial records, health information), it’s not. The answer is self-hosting — which DeepSeek’s open weights make possible, but which requires technical capability most NZ businesses don’t have yet.
❓ Frequently Asked Questions
Q: Is DeepSeek actually as good as GPT-5.5 or Claude Opus? On most standard benchmarks, DeepSeek V4 Pro scores within 3-5% of GPT-5.5 and Claude Opus 4. It falls behind on some reasoning tasks and English-language creative writing. For most production workloads — code generation, analysis, summarisation, translation — the performance gap is negligible.
Q: Why would DeepSeek make a discount permanent? Two likely reasons: the cost structure actually supports it (Chinese compute is cheaper), and permanent low pricing is a moat. If DeepSeek is always 75% cheaper, switching costs become real. Every month a developer stays on DeepSeek is a month they’re not building OpenAI into their stack.
Q: What are the data sovereignty risks for NZ businesses? DeepSeek’s API processes data on servers subject to Chinese law, including data access provisions that differ significantly from NZ’s Privacy Act. For non-sensitive workloads (content generation, code assistance), this may be acceptable. For anything involving personal data, financial records, or health information, self-hosting on NZ infrastructure is the only compliant option.
Q: Will Western AI companies match the price? They can’t without destroying margins. OpenAI’s latest funding round valued the company at $300 billion. That valuation assumes revenue growth that depends on premium pricing. Anthropic is in the same boat. Google could subsidise Gemini pricing from search revenue, but it would cannibalise its own cloud margins. The structural cost advantage is real.
🔍 THE BOTTOM LINE
DeepSeek’s permanent 75% discount isn’t a promotion — it’s a declaration. Chinese AI can match Western capability at a quarter of the price, and it can do it sustainably. The AI industry’s pricing assumptions just collapsed. Every board room at OpenAI, Anthropic, and Google is having the same conversation this morning: cut margins or lose developers. There isn’t a third option.
SOURCES
- Bloomberg — DeepSeek to make permanent 75% discount on flagship AI model
- Hacker News — 42 points, 41 comments in first hour
