DeepSeek open sources its agent software: Harness and time of day API pricing
DeepSeek released Harness v0.1, an agent platform under the MIT license, and announced its new model deepseek-v4-pro-0813. On pricing there is an unusual choice: the API rate changes with the time of day and the gap between the two tiers is twofold.
Harness: an agent platform where every component swaps out
Per a report from The Decoder dated August 13, 2026, DeepSeek Harness v0.1 is an agent platform released under the MIT license.
The distinguishing part of the architecture is its modular plugin structure. In the words of the report, all features are swappable plugins, from tools and sandboxes to sessions and the UI.
On the model side the new version is named deepseek-v4-pro-0813. The company states that existing integrations will keep running without any tweaks, so the transition is backward compatible.
A Webrazzi report dated August 14, 2026 covers the same development and frames Harness as an open source rival to Claude Code.
API pricing that changes by the hour
Pricing takes effect on August 16 and comes in two tiers. Off peak it is $0.66 per million input tokens and $1.98 per million output tokens.
At peak the price doubles exactly: $1.32 per million input tokens and $3.96 per million output tokens.
Cache hits split the same way: $0.022 per million tokens off peak and $0.044 at peak.
Per the report the rate hours are arranged around business hours in China. On performance, Artificial Analysis data shows V4-Pro rising from 45 to 53 on the Intelligence Index while remaining behind Claude Opus 5 at 63 points and Kimi K3 at 60.
Tying price to the hour is a capacity signal
The most interesting choice here is the tariff rather than the model. An API price doubling by time of day means a logic familiar from electricity and cloud capacity arriving in AI: demand is straining the peak, and price is being used to spread it. That assessment is ours.
Second, this creates a concrete saving on the business side. Work that does not need to be real time, such as batch content generation running overnight, data classification or report summarization, can shift to the cheap tier. Same job, same model, half the price.
Third, the relationship between open source and price. Giving the agent software away under MIT makes the software free and the model paid. That is a familiar strategy of spreading the infrastructure layer while moving revenue to the model. The source does not interpret it this way, the inference is ours.
Fourth, its place in the benchmarks. Rising from 45 to 53 on the index is clear progress, and the report states plainly that it stays behind the leader. The cheaper model may not suffice for every task, and the decision sits in the balance between accuracy and cost.
What it means for businesses in Türkiye
The assessment below is not in the sources, it is our reading. The reports contain no Türkiye specific information.
The most practical takeaway for businesses here concerns timing. Time of day pricing turns the scheduling of non real time workloads into a direct cost decision. For an ecommerce site generating thousands of product descriptions, doing it overnight rather than during the day creates a twofold difference.
The second point is data. Using the model means data going to the provider. For teams working with sensitive customer data that is a question to ask before the price advantage. Open weight models that run on device offer an alternative here. That connection is our note.
Third, the agent software being open source. For a team wanting to build its own infrastructure, an MIT licensed platform is a serious shortcut compared with writing from scratch. But open source does not mean free: the setup, maintenance and security burden stays with you.
The practical suggestion: split your workloads into real time and batch. Shift the batch ones to the cheap tier, and for the real time ones choose the model by required accuracy. Those two steps produce measurable savings in most accounts.
The UNALSOFT take
In the flows we build on the agentic AI side, the most efficient step in cost optimization is almost never changing the model, it is placing the workload at the right time and on the right model. Time of day pricing makes that visible again: running the same job at a different hour means half the cost without changing anything else. That is why the first thing we do with clients is split workloads into real time and batch. An AI budget run without that split always costs more than it needs to.
Sources
The Decoder, DeepSeek V4-Pro and open source agent software · Webrazzi, DeepSeek releases Harness
Have you split your AI workloads?
Let us separate real time from batch work and bring the cost down together.