DeepSeek has rolled out V4 Pro, the newest version of its flagship artificial intelligence model, while simultaneously announcing a restructured pricing regime that will see API costs climb sharply from August 17.
The Chinese AI company is positioning the update around improved agent capabilities, with the V4-Pro-0813 release now accessible through its API, app, and web platforms. But the accompanying price adjustments, which introduce separate peak and off-peak rates, mean developers and businesses relying on DeepSeek’s models will need to recalculate their budgets.
Depending on the specific model and token type, increases will range from 50 percent to as much as 1,100 percent. The shift moves DeepSeek away from the aggressive low-cost positioning that helped it capture global attention and toward a pricing structure that reflects the actual expense of running large-scale AI infrastructure.
At current rates, V4 Pro charges $0.435 per million input tokens when the cache misses, while output runs at $0.87 per million tokens. The model supports a context window of one million tokens, a capacity that sets it apart from many competitors.
The underlying architecture remains consistent with the V4 family introduced in April. V4 Pro employs a mixture-of-experts design containing 1.6 trillion total parameters, though only 49 billion activate for any given token. DeepSeek has also leaned on compressed attention techniques to keep computational demands manageable, claiming the model requires roughly 27 percent of the single-token inference computing that V3.2 needed at the same one-million-token context length.
Agent performance is the headline story. DeepSeek reports a score of 87.9 on Terminal-Bench 2.1 and 83.3 on CyberGym, both significant gains over previous versions. These are internal benchmarks, so independent evaluation will determine how those numbers hold up in real-world use. Artificial Analysis has already positioned the V4 Pro family among the stronger open-weight models, assigning the reasoning variant a score of 44 on its Intelligence Index in the most recent assessment.
The broader context of DeepSeek’s journey matters here. When R1 landed in early 2025, it proved that a Chinese firm could build a competitive reasoning model without matching the capital outlays of the largest American AI labs. That narrative has shifted. Rivals such as Moonshot AI, Zhipu, MiniMax, Alibaba, and ByteDance have all released capable models, forcing DeepSeek into a continuous cycle of product iteration and infrastructure investment.
The financial stakes are rising accordingly. DeepSeek secured approximately $7.4 billion in funding in June and is reportedly exploring another round that could value the company near $74 billion. That would mark its second external raise in quick succession after years of relying on founder Liang Wenfeng’s quantitative investment firm, High-Flyer, for backing.
The money is flowing into more than model research. DeepSeek is expanding its workforce across data centre operations and AI agent teams while accelerating hiring of chip-design engineers. The goal is an in-house inference chip that could reduce reliance on suppliers such as Nvidia and Huawei, a strategic priority for Chinese technology companies navigating restrictions on advanced foreign hardware.
For Nigerian developers building on DeepSeek’s API, the price increases will force hard choices. Applications that once ran on tight budgets may need token optimisation, caching strategies, or migration to alternative models. The peak and off-peak pricing structure adds another layer of complexity, rewarding users who can schedule workloads strategically.
The V4 Pro launch is therefore not simply a product update. It signals DeepSeek’s transition from a scrappy disruptor to a heavyweight contender spending billions on people, computing infrastructure, and silicon while asking customers to share more of the cost.
