DeepSeek's V4-Pro-0813 ships with better agents and peak-hour API pricing

DeepSeek released V4-Pro-0813 on August 13 and said the model substantially improves the capabilities of AI agents, systems that carry out complex multi-step tasks on their own ^[1][2][4]. The part that will show up on an invoice is separate: the company plans to raise API prices for V4 Pro and V4 Flash and to charge according to load, with peak-demand hours priced higher and quieter periods lower ^[5][6].

The model is live through the API, the app and the web version ^[3]. The pricing change, as reported by Reuters and summarised by mezha.net, arrives without the details a buyer needs to plan against it: no rate card, no definition of which hours count as peak, no effective date ^[18].

Time-of-day pricing is a familiar utility mechanic, and its purpose is to push flexible demand out of the busy window. For a buyer it converts a fixed unit cost into a scheduling problem. Batch work, evals, backfills, bulk document extraction, can move to the cheap hours. Interactive traffic cannot, and neither can an agent run triggered by a user in the middle of the working day. That is the uncomfortable overlap here: the capability DeepSeek is advertising is the one whose call volume per task is hardest to forecast, because a multi-step agent determines how many steps it takes ^[4], and it will be used at exactly the hours the meter is set to run fastest.

There is a second reason to benchmark before committing. In independent testing, the cheaper V4 Flash unexpectedly outperformed the earlier V4 Pro version DeepSeek showed in April ^[7], an inversion of the company's own positioning, in which Pro is the more powerful product ^[8]. DeepSeek also shipped V4 Flash as the cheapest known global model to run, with test costs far below rivals though quality behind the leading competitors ^[9][10]. Whether the official Pro release has closed that gap is the first thing to re-run internally, because it decides which model you point peak-hour traffic at.

The commercial pressure behind all of this is visible. DeepSeek raised roughly $7.4 billion in its first external round in June, and in July was reported to be planning another at a valuation near $74 billion, about ten times the June raise ^[13][14][15]. It intends to at least double headcount across units including the data centre and AI agent teams ^[16], and has stepped up non-public hiring of chip design engineers to build its own AI processor, which would cut dependence on suppliers including Nvidia and Huawei ^[17]. Its early lead in China is under pressure from products by Moonshot AI, Zhipu AI, MiniMax, Alibaba and ByteDance ^[12], after R1 made it one of the country's most visible AI companies in early 2025 ^[11].

Watch for the published rate card and the definition of peak, and whether the tiering applies to the app and web products or only to the API ^[5][6][3]. Watch whether independent tests place the final V4 Pro above V4 Flash ^[7]. And watch whether the reported round closes at the reported valuation ^[14]; a company doubling its data centre staff and designing silicon ^[16][17] has reasons to want peak-hour revenue that have nothing to do with how good the agents turn out to be.