DeepSeek opens up coding agent Harness and updates V4-Pro

What's new

Chinese AI company DeepSeek launched two things on August 13, 2026: DeepSeek-V4-Pro-0813, the official version of the company's flagship model, and DeepSeek Harness v0.1. Harness is what is known as an agent framework, a program that lets an AI model work on tasks on its own. It can edit files, run commands in the terminal, make a plan and use subagents, which handle subtasks.

Harness is open source under the MIT license and is hosted on GitHub, so anyone can download and modify the code. It is only a developer preview (version 0.1), however, so it is aimed at developers and is not ready for everyone. According to The Decoder, 712 projects signed up as testers within three days.

DeepSeek has also launched a Search/Fetch API that can search the web and retrieve pages through the company's own search engine.

What's smart

Harness is built modularly on the principle that everything is a plugin. The architecture is called Cordis and makes tools, sandboxes (protected environments where code can run without breaking anything) and user interfaces interchangeable. It supports the OpenAI Responses API and has adjustable levels for how thoroughly the model thinks.

DeepSeek's API can also be used as the engine in Claude Code, GitHub Copilot and OpenCode without extra coding, and it understands both OpenAI's and Anthropic's formats. That means you can try DeepSeek's model in tools you already know.

Cheaper or better?

Better: According to DeepSeek's own figures, V4-Pro-0813 rose on Terminal Bench 2.1 from 72.1 to 87.9, and on DeepSWE from 12.8 to 62.7. It also scored 74.1 on Toolathlon-Verified and 71.1 on DSBench-FullStack. The figures have not been independently verified. On the Artificial Analysis Intelligence Index, V4-Pro scored 53 points, compared with 63 for Claude Opus 5 and 60 for Kimi K3. So the model is not at the top.

Cheaper: No, quite the opposite. From August 16, DeepSeek has introduced peak and off-peak pricing. The prices below are per million tokens (chunks of text) in dollars.

V4-Pro previously cost 0.435 for input and 0.87 for output. Now the off-peak price is 0.66 and 1.98, and the peak price is 1.32 and 3.96. V4-Flash previously cost 0.14 and 0.28. Now the off-peak price is 0.22 and 0.66, and the peak price is 0.44 and 1.32. Cached input on V4-Pro (reused text, which is cheaper) rose from 0.003625 to 0.022 off-peak and 0.044 at peak.

According to The Decoder, the pricing periods follow Chinese business hours, and the change partly reverses discounts from May. Monthly prices for ordinary users are not stated in the sources.

What it's good for

Harness and V4-Pro are aimed at coding and agent tasks, where the AI writes, tests and fixes code itself. VentureBeat names Anthropic's Claude Code and OpenAI's Codex as the closest competitors. For developers and companies with many automated tasks, it may be worth testing, but the price increase has to be factored in.

The API is available at api.deepseek.com and requires an API key. The documentation does not specifically mention availability in Denmark or the EU, and it is unclear how data processing is set up here. That is worth clarifying before company data is sent over.