DeepSeek Is Raising API Prices by Up to 1,100% Ahead of a Possible IPO

DeepSeek is raising API prices by up to 1,100% for peak hours, with Tencent Cloud's June 29 notice showing V4-Pro and V4-Flash rates doubling during 09:00-12:00 and 14:00-18:00 Beijing time, ahead of a possible mainland China IPO. Bloomberg reported in July that DeepSeek is seeking private funding at a valuation of at least 480 billion yuan ($71 billion) and could file for an IPO as soon as this year, with a market debut possible in 2027. The price hike, while still undercutting OpenAI and Anthropic, signals a shift from DeepSeek's low-cost positioning as it prepares for public investors.

DeepSeek is changing the cheap API story that made developers pay attention. The price gap with OpenAI and Anthropic is still there, but it no longer looks as effortless as it did a few months ago. DeepSeek's V4 pricing shift is a small table with a large message inside it. Tencent Cloud said in a June 29 notice that DeepSeek V4-Pro and V4-Flash would move to peak and off-peak billing around the V4 official release, with peak rates set for 09:00 to 12:00 and 14:00 to 18:00 Beijing time. During those windows, the listed rates double. That's not a footnote. It's the bill. For V4-Pro, Tencent's notice listed regular pricing at 0.025 yuan per million cache-hit input tokens, 3 yuan per million cache-miss input tokens, and 6 yuan per million output tokens. Peak pricing rises to 0.05 yuan, 6 yuan, and 12 yuan for the same three items. Double, across the board. For V4-Flash, regular output is 2 yuan per million tokens and peak output is 4 yuan. The company's own API pricing page still shows the U.S. dollar list price for V4-Pro output at $0.87 per million tokens and V4-Flash output at $0.28, with lower cache-hit input rates that make repeated prompts far cheaper than fresh context. You can see why developers liked this. A product that sends long prompts, tool definitions, memory, retrieval snippets, and agent instructions into a model all day can burn money fast. DeepSeek made that feel less frightening. Cache hits were cheap. Output was cheap. The whole pitch was simple: close enough on quality, much cheaper on use. That pitch still works. It just has a clock attached to it now. The cheap model now has business hours Peak pricing is not the same as a clean across-the-board price rise. If your workload runs overnight, you can schedule around it. If your app serves office-hour traffic in China, you can't pretend the higher line in the table is theoretical. Live chat, coding agents, support bots, and internal copilots tend to be used when people are working. That means the peak window is exactly when many real customers show up. Here's the thing: DeepSeek's advantage has never been that it matched every frontier model point for point. It was that a company could test a capable model without asking finance to approve a Silicon Valley-sized inference bill. That matters when you're routing thousands of small decisions through an AI layer, not running one impressive demo for investors. The cache line is especially important. DeepSeek's docs say cache hits are charged far below cache misses, and that is where agent workflows can win or lose. A stable system prompt, repeated tool schema, or reused document prefix can keep costs down. A messy workflow that keeps missing the cache pays the larger input rate. The model may be cheap, but bad architecture still sends you a bigger invoice. Don't ignore that part. The IPO story makes the pricing easier to read Bloomberg reported in July, in an article carried by Investing.com, that DeepSeek was preparing for a possible mainland China IPO filing as soon as this year and was seeking fresh private funding at a valuation of at least 480 billion yuan, or about $71 billion. The same report said a filing could lead to a market debut in 2027. Once you put that beside the pricing change, the move looks less mysterious. Underwriters don't sell a charity project. Public-market investors want proof it isn't one either. They want revenue quality, margin, and some proof that a popular model can become a durable business, not just a lab whose commercial pitch is inference for next to nothing, forever. DeepSeek is still undercutting the big American labs by a wide margin on published API prices. That's real. OpenAI and Anthropic remain far more expensive at the top end of their model lines, especially for output-heavy work. But the clean old story, DeepSeek as the permanent low-cost spoiler, is getting more complicated: a company preparing for public investors has different incentives from one trying to shock the market into adoption. That doesn't make DeepSeek less dangerous to its rivals. It may make it more normal. The strongest startups often start by making the incumbent's pricing look absurd, then raise their own prices once users have built around them. You don't need a theory for that. You need only look at the bill after the promotion ends. For developers, the practical answer is plain. Keep using DeepSeek where the quality and cost make sense, but model peak traffic instead of quoting the lowest line on the pricing page. Watch cache-hit rates. Separate batch work from live work. And don't build a business case around the idea that any AI provider will stay cheap just because it started cheap. DeepSeek's discount story isn't over. But it is becoming a business story, and business stories eventually find their way into pricing tables. Also read: Anthropic Found Its Claude Agents Attacked Each Other With Malware https://startupfortune.com/anthropic-found-its-claude-agents-attacked-each-other-with-malware/ • Lam Research Stock Jumps 4.7% to Push Its Market Cap Past $400 Billion https://startupfortune.com/lam-research-stock-jumps-47-to-push-its-market-cap-past-400-billion/ • Samsung Sold Out Every Gigabit of Its 2026 HBM4 Memory Chips https://startupfortune.com/samsung-sold-out-every-gigabit-of-its-2026-hbm4-memory-chips/