DeepSeek makes V4 Pro generally available and rebuilds its price list around time-of-day tiers

DeepSeek has announced the general availability of DeepSeek-V4-Pro-0813, replacing the preview version of its flagship model. According to the technical documentation, the model has 1.7 trillion parameters, an MIT licence and a context window of one million tokens, alongside a new speculative decoding module called DSpark. The new features include the `reasoning_effort` parameter, offered at three levels to modulate processing, and native support for OpenAI's Responses API. The agentic benchmark results listed in the model card, such as Terminal Bench 2.1 and DeepSWE, remain for now figures declared by the company and awaiting independent validation.

The change with the greatest practical impact concerns the economics of the service: from 16:00 UTC on 16 August, DeepSeek is dropping its single price list in favour of two fixed-price time bands, peak and off-peak. The documentation lists peak hours as the 01:00–04:00 and 06:00–10:00 UTC windows, with off-peak rates set at half the peak ones. Comparing the old and new price lists for V4 Pro shows a marked rise: peak-hour output goes from 0.87 dollars to 3.96 dollars per million tokens, while cache-hit input records the steepest percentage increase, climbing roughly twelvefold to 0.044 dollars. Even off-peak, the list stays above the previous one: output moves from 0.87 to 1.98 dollars per million tokens. The decision marks a reversal of the aggressive pricing policy that had defined the company in the open-model space.

The release lands amid industrial expansion. As reported by Reuters, DeepSeek raised roughly 7.4 billion dollars in June 2026 in its first external funding round, and is said to be weighing a new round at a valuation of around 74 billion — a figure that nonetheless remains an agency report, not officially confirmed. All of this is happening while domestic Chinese competition — from Alibaba and ByteDance to MiniMax, Moonshot AI and Zhipu AI — steps up the pace of releases. The MIT licence keeps the weights in the open-release category, but the cost of using the model through the API no longer depends only on which model you pick: it also depends on what time you query it.

Gaps remain in the available information. Although the time bands are expressed in UTC, the source does not specify whether they derive from a conversion from Beijing time, leaving the practical impact on European users uncertain until the change takes effect. On top of that, the absence of an official price-performance comparison with competing models, and the lack of architectural detail beyond the DSpark module, limit any full assessment of how the product sits in the market.

— Olya