NUPIAO Reporter | Song Jianan
Mark your calendars! On June 29, DeepSeek dropped an upgrade email to its users confirming that the full-fledged DeepSeek V4 is officially scheduled for launch in mid-July. We’re talking serious optimizations and performance boosts coming your way.
Here’s the catch—and it’s important: To keep things running smoothly and allocate resources wisely, DeepSeek announced they’ll be tweaking their API pricing strategy once the正式版 (official version) drops. They’re introducing a peak-and-valley pricing model. What does that mean for you? If you call the API during peak hours, expect the price to literally double.
Let’s break down the numbers so you know exactly what to budget:
- DeepSeek V4 Pro: For 1 million input tokens (with cache hits), the usual rate is 0.025 yuan, but during peak hours, it jumps to 0.05 yuan. If the cache misses, it goes from 3 yuan up to 6 yuan. For outputs, you’re looking at 6 yuan normally, rising to 12 yuan during rushes.
- Peak Times Defined: Beijing Time 9:00–12:00 and 14:00–18:00 are now considered “rush hour.”
Don’t worry if you’re on a tighter budget; DeepSeek V4 Flash remains more affordable. Its 1 million input tokens (cache hit) cost 0.02 yuan normally, jumping to 0.04 yuan at peak. Cache misses go from 1 yuan to 2 yuan, and outputs rise from 2 yuan to 4 yuan.
NUPIAO wants to make sure you’re in the loop: The team promises to send out a 24-hour advance notice via email before any price changes take effect. If you stick with DeepSeek after the adjustment, it counts as your agreement to the new rates. Not happy with the change? You can opt out and request a refund.
For context, the V4 Preview launched back on April 24th with both Pro and Flash versions. Both boast a massive 1-million-token context window and support enterprise-grade features like switching thinking modes, JSON output, tool calls, and prefix-based writing. It’s built to handle complex scenarios in dev, office work, law, and finance.
Compared to previous models, the Agent capabilities have gotten a serious upgrade. Reports say V4 Preview is already the go-to Agentic Coding model for DeepSeek’s own internal team. Early feedback suggests the user experience beats Sonnet 4.5, and the delivery quality is nearly on par with Claude Opus 4.6 in non-thinking mode (though there’s still a gap compared to Opus 4.6’s thinking mode).
While V4 Flash might not have quite the same world knowledge base as the Pro version, its reasoning skills are surprisingly close. Thanks to smaller parameters and activation requirements, V4 Flash delivers a faster, more wallet-friendly API service.
In agent benchmarks, V4 Flash holds its own against Pro on simple tasks but still trails off when things get really complicated.
Remember those crazy low prices from the preview phase? Flash was 0.2 yuan per million tokens for cache hits, while Pro was a steep 1 yuan. Now, thanks to the upcoming availability of Ascend super-node products later this year, DeepSeek hints that Pro version prices could drop significantly, making high-performance AI even more accessible.
Speaking of prices, back on April 26th, DeepSeek made headlines by slashing API costs across the board. Input cache hit prices dropped to one-tenth of the launch price. V4 Pro got an extra 75% discount (2.5折), pushing the cost for 1 million cached inputs down to a record-breaking global low of just 0.025 yuan.
According to the official pricing page, the last round of cuts focused heavily on cache-hit scenarios. V4 Flash saw its cache-hit input price plummet from 0.2 yuan to 0.02 yuan per million tokens.
The Enterprise Pro version got the biggest love, dropping from 1 yuan to 0.1 yuan for cached inputs. Before May 5, 2026, there was a limited-time 2.5折 deal, making it just 0.025 yuan. Uncached inputs fell from 12 to 3 yuan, and outputs from 24 to 6 yuan. Just remember, some of these were temporary promos.
Analysts tell us this isn’t just a random price hike. In a world where compute power is scarce, the peak-valley pricing is actually a smart standardization tool. Data from OpenRouter shows that V4 Flash alone has handled over 4.66 trillion tokens in a single week, topping global charts for six weeks straight—though usage did dip slightly by 6% recently. The real issue? Office hours are clogged, leading to frequent timeouts and service lag.
We’ve seen DeepSeek face stability issues before because the rock-bottom prices attracted too much traffic, pushing their compute clusters to the breaking point. This time, raising prices during peak hours acts as a lever to push non-urgent, batch-processing tasks to off-peak times. This ensures that high-priority stuff like financial trading, code development, and real-time agents stays stable when people actually need it most.
Before the official release, DeepSeek already rolled out a big update called DSpark for the preview version. It’s a speculative decoding framework that boosts inference speed by 60% to 85%. While currently for the preview, it’s a strong hint that the final version will be even more efficient and cost-effective.
And guess what? The money talks are heating up too. Rumors from June 16 suggest DeepSeek just closed its first external funding round, raising over 50 billion yuan and valuing the company at more than 338 billion yuan post-money.
Insiders reveal that founder Liang Wenfeng might have invested about 20 billion yuan himself, making him the largest single investor. Tencent chipped in around 10 billion, while the CATL ecosystem (including CATL and Puquan Capital) put in roughly 5 billion. NetEase, JD.com, Monolith, IDG Capital, Zhengxin Valley, and Shixiang Technology also joined the party with investments ranging from 15 to 30 billion yuan each.
Hiring is going wild as well. DeepSeek is planning to double the size of every department. They’ve opened 33 positions across algorithms, R&D, operations, product, data engineering, and admin roles in Beijing and Hangzhou—and yes, they’re taking interns!
Looking ahead, the official V4 release will plug the commercial gaps left by the preview. With a mountain of capital and talent coming on board, DeepSeek is poised to close the gap with top-tier overseas closed-source models in terms of real-world business adoption.
