DeepSeek has done something unusual for a frontier-class model provider. It is keeping the discount. The Chinese lab confirmed this week that the 75 percent reduction it had been running on its flagship V4-Pro API will not roll back at the end of May. The sale becomes the standard. V4-Pro will now sit at about $0.435 per million input tokens and $0.87 per million output tokens, with cached input at a strikingly low $0.003625 per million. The cheaper V4-Flash variant comes in at $0.14 in and $0.28 out. For anyone who has built a serious product on top of OpenAI or Anthropic pricing this year, the spreadsheet just changed.
What makes this more than a competitive nudge is the scope. DeepSeek did not just trim its top-tier model. It cut cache-hit prices across the entire lineup to one-tenth of launch pricing, dating that change back to late April. That detail matters because the cache hit is where high-volume AI products actually live. Coding assistants, customer support agents, research tools and document workflows pass the same system prompts and reference material through the model again and again. A tenfold reduction on cached input is, for those products, a much bigger structural change than the headline 75 percent.
For startups, the new pricing rewrites the basic margin question. AI-native products often look like software, but their cost structure has behaved more like usage-based infrastructure: every answer, every report, every autonomous step generates a real token bill. Founders have spent the past year either keeping prompts brutally trimmed or watching gross margin disappear. DeepSeek's reset opens room to keep more context in the model, run more steps, or serve markets where users will not absorb enterprise pricing. It also lowers the threshold at which a small team can credibly offer an agentic product without immediate burn.
The strategic logic is less comfortable. DeepSeek is compressing its own economics at exactly the moment frontier AI costs remain enormous. According to Bloomberg, the company is in the middle of a roughly 70 billion yuan funding round, around $10 billion, and has been telling potential investors it will prioritise breakthrough research over near-term revenue. Founder Liang Wenfeng has also reportedly committed to continuing open-source releases while chasing artificial general intelligence. Cutting prices, staying open, and underwriting a research roadmap that takes years to mature is a bold combination. It works if DeepSeek's architecture really is more efficient, if the funding round closes, and if developer adoption translates into platform usage that survives geopolitical friction.
Geopolitics is the part that prevents this from being a clean rerun of how Amazon Web Services priced rivals out of object storage. Enterprise buyers in the United States and Europe are wary of building sensitive workloads on a Chinese provider, and Washington's stance on Chinese AI exports is unsettled even as the Trump administration negotiates with Beijing. Many teams will treat DeepSeek as a benchmark and a fallback rather than the default route, particularly anywhere data residency or procurement reviews are involved. That does not blunt the pressure on rival pricing. It just channels it through architecture, with developers building routers that send the cheap, repetitive work to DeepSeek and reserve premium models for the hardest tasks.
The harder question now sits with OpenAI, Anthropic and Google. None of them needs to lose every customer to feel this; they just need procurement teams to start asking why a similar workload costs several times more elsewhere. The frontier labs have spent the past two years arguing that quality and reliability justified a wide price premium. DeepSeek is testing whether they were right. If V4-Pro and V4-Flash hold up under sustained production traffic at these new rates, the era of competing primarily on benchmark scores is closing. The next phase of the market will be fought on the daily economics of products that have to make money, and the AI labs that priced like the technology was scarce are running out of months to explain why.