The number arrived

DeepSeek's new interface prices take effect at 16:00 UTC on August 16. On V4-Pro, cache-miss input moves from 0.435 dollars to 0.66 dollars per million tokens and to 1.32 dollars at peak; output rises from 0.87 dollars to 1.98 dollars and 3.96 dollars. The sharpest rise lands on the cache hit: from 0.003625 dollars to 0.022 dollars, and to 0.044 dollars at peak. The peak definition is not arbitrary either; the 01:00–04:00 and 06:00–10:00 UTC windows line up with the Chinese working day.[1]

On August 6 this column recorded that the announcement carried neither a schedule nor a figure, and that the whole planning burden sat with the buyer. The schedule is now published, and where the burden lands is visible: the bill turns on what hour a request is sent. Tying a cost line to a tariff does not have to make the buyer's job easier; what it does make easier is the seller's job of spreading demand across the day.[1], [3]

The harness opened on the same day

In the same announcement the company published its Node.js coding harness under an MIT licence. Every part of it is a replaceable plugin, including the model adapter, the tool registry, the session log and the agent loop itself, and the documentation states there is no privileged core to patch. The provider catalogue covers Anthropic, OpenAI, AWS Bedrock, Microsoft Azure and Google, with room for custom OpenAI-compatible gateways on top. The repository picked up more than 33,000 stars within a few hours.[2]

Read together, the two decisions draw a line of ownership. The layer handed to the developer is the one that can be argued over and swapped out; the layer DeepSeek keeps is the endpoint whose price it sets itself. That the harness also supports rival providers does not erase the line, it only lowers the cost of moving. There is another reading of the same facts: with rival harnesses multiplying, opening your own can be a developer-relations move with no connection to pricing. What would tell the two apart is which endpoint the people using the open harness actually send their requests to.[2], [1]

The part a number can settle

It is also worth noting which line grew most in proportional terms: the cache hit. A cache hit is the discount that applies when the same context is sent again and again, and agent loops that carry the same history through long sessions lean on exactly that discount. The price schedule narrows it, while the newly opened harness keeps a continuous session log and lets runs be replayed. The two decisions do not point the same way. Even so, that alone is not proof of targeting; the company may simply be repricing what its own cache infrastructure costs it.[1], [2]

A number is what will settle this. If DeepSeek publishes token volumes for peak and off-peak hours after August 16, or an average realised unit price, whose ledger the increase lands in becomes measurable. If it does not, the average unit price that independent interface-cost trackers report for V4-Pro answers the same question approximately. That is the signal to watch; what has been published so far is only the tariff.[1]