DeepSeek Open-Sources Agent Harness While Hiking API Prices Up to 12x
On the evening of August 13, DeepSeek dropped two pieces of news at once. Developer groups were simultaneously sharing Harness code and calculating API bills—how much did prices actually go up? The mood was quite mixed.
First, the official version of V4 Pro went live. The industry had already heard rumors about this, but what really made people restless was the accompanying API price adjustment announcement. DeepSeek had been brewing this price increase for a long time. The official website put up a notice early on, clearly stating four words: significant increase expected. When the prices actually came out, developers found those two words were no exaggeration.
Don't rush to worry about the price yet; let's first look at this heavyweight tool!
DeepSeek Harness Developer Preview is open for global Harness developers to test, with source code released simultaneously.
All Agent capabilities—models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI—are composed of plugin combinations, freely replaceable and flexibly reorganized.
Harness: A Set of "Building Block Tools" for Agent Developers
DeepSeek Harness was the news released that evening, Developer Preview v0.1, MIT license, code released directly. Unlike previous model weight releases, what was open-sourced this time was the entire set of Agent runtime infrastructure outside the model.
Its design philosophy is called "everything is a plugin." Models, tools, skills, sessions, sandboxes, storage, Agent loops, task scheduling, even the UI can be loaded, unloaded, and replaced as plugins. Want to change models? Swap the model plugin. Want to change scheduling? Swap the scheduling plugin. The degree of freedom is substantial.
Getting started is also quick. Just install Node.js and start the service with one command:
npx @deepseek-ai/dsh web
After running, open http://127.0.0.1:3080/ in your browser, and the local Web UI is ready to play with.
If you want to dig deeper, pull the source code:
git clone https://github.com/deepseek-ai/deepseek-harness
Harness comes with four built-in running modes:
- Standard mode provides the full set of tools
- PTC mode lets you use model-generated code to combine multiple rounds of tool calls
- Minimal mode leaves only a shell and a file editing tool, specifically for running benchmarks in minimal environments
- Creative mode allows experimenting with plugins at runtime and assembling new playstyles yourself
Of course, custom model integration is also supported:
The official statement also noted that this is only an early preview version and will iterate rapidly in the future. Much anticipation.
Price Adjustment: Peak-Valley Pricing Arrives, Up to 12x Increase
The other half of the news on the same day was the comprehensive API price adjustment. The core is the introduction of peak-valley pricing: off-peak prices are half of peak prices. Peak hours are Beijing time 9:00-12:00 and 14:00-18:00; the rest is considered off-peak. The new prices take effect at 00:00 on August 17.
For specific numbers, putting the prices before and after the increase side by side is more intuitive. The units below are all per million tokens, divided into pre-adjustment, off-peak, and peak tiers. Pre-adjustment prices are back-calculated based on the officially announced increase multiples:
| Model | Billing Item | Pre-Adjustment | Off-Peak | Peak |
|---|---|---|---|---|
| V4 Flash | Input (Cache Hit) | 0.02 CNY | 0.05 CNY | 0.10 CNY |
| V4 Flash | Input (Cache Miss) | 1.0 CNY | 1.5 CNY | 3.0 CNY |
| V4 Flash | Output | 2.0 CNY | 4.5 CNY | 9.0 CNY |
| V4 Pro | Input (Cache Hit) | 0.025 CNY | 0.15 CNY | 0.30 CNY |
| V4 Pro | Input (Cache Miss) | 3.0 CNY | 4.5 CNY | 9.0 CNY |
| V4 Pro | Output | 6.0 CNY | 13.5 CNY | 27.0 CNY |
At a glance, it's clear whether prices went up: V4 Pro peak output is 27 CNY, 4.5 times the pre-adjustment price of 6 CNY, an increase of about 350%; V4 Flash peak output is 9 CNY, 4.5 times the original price of 2 CNY. The cache hit tier saw the steepest increase, with Pro jumping from 0.025 CNY straight to 0.30 CNY—a 12x increase.
Looking at Pro overall, the new prices range from 1.5x to 12x the original, with increases falling in the 0.5x to 11x range. The highest single item, peak-hour cost, is roughly 4.5 times the pre-adjustment price. For a team running business on APIs, this is no longer a "slight increase" concept.
One point worth noting: the bulk of expenses comes from one-off tasks and high-frequency polling, while cache hits can save a lot. Taking V4 Flash peak hours, a hit saves about 96% compared to a miss. Many repetitive prompts can be handled entirely by caching to keep costs down.
Is This a Coincidence?
Why release both on the same night? On the surface, these two matters seem opposite—one free, one charging—but it's actually a combination play. Open-sourcing Harness builds the ecosystem, bringing more developers in to build their own Agents, which naturally consumes more tokens.
Meanwhile, the API price increase supplements commercialization, shifting computing power costs onto truly heavy usage. Peak-valley pricing is also shrewd, guiding everyone to move non-urgent batch processing tasks to off-peak hours, easing peak computing pressure.
For ordinary developers, it boils down to the following three points:
- First, move non-real-time tasks to off-peak hours; it's half the price.
- Second, use caching whenever possible; the price difference for cache hits is nearly 20x.
- Third, if you're just trying things out, run Harness locally with
npx @deepseek-ai/dsh webfirst, swap models for fun, and keep the bill stable for now.
DeepSeek's move this time is less about raising prices and more about re-pricing the entire ecosystem. Tools are given away for free, computing power is clearly priced. Now it's a matter of seeing who runs ahead first.
As for DeepSeek Harness's "everything is a plugin"—whether it's good or not, we'll wait and see!
Related Resources
- DeepSeek Harness Official Developer Site:
https://deepseek.com/harness - GitHub Repository:
https://github.com/deepseek-ai/deepseek-harness - Quick Experience:
npx @deepseek-ai/dsh web, openhttp://127.0.0.1:3080/in browser