Updates·August 14, 2026, 15:36

DeepSeek opens up coding tool it positions as an alternative to Claude Code

AI-generated from the sources listed, with automated source checks.

Chinese AI company DeepSeek has released Harness v0.1, an open tool designed to let the company's models write code on their own. The final V4 Pro also launched, and API pricing changes on August 16, according to DeepSeek's plan.

DeepSeek released Harness v0.1 on GitHub on August 13, 2026. It is an early version, called a developer preview, and is therefore not production-ready, according to Crypto Briefing.

Harness is what is known as an agent runtime framework. That means a kind of toolbox meant to turn DeepSeek's V4 models into autonomous coding agents. An agent is an AI program that can take several steps in a row on its own and use tools, instead of simply answering one question at a time.

Open code and free use

Harness is open source under the MIT license, which permits commercial use. It is built on the meta-framework Cordis and can work with codebases, terminals and external tools. According to Forklog, DeepSeek positions it as a minimalist framework that can also be used with proprietary models, and as an open alternative to Anthropic's Claude Code.

The Harness team is led by Cui Tianyi, according to Crypto Briefing. He previously worked at Jane Street and joined DeepSeek in March 2026.

The lead-up

In May 2026, DeepSeek announced that a new Harness team in Beijing would build a coding agent called DeepSeek Code. Researcher Deli Chen announced it on X, and job postings were put up for a product manager and a developer. The postings required experience with agent loops, MCP, multi-agent systems and context engineering.

According to PYMNTS' account of a Bloomberg report from August 12, DeepSeek had also created a WeChat account for "DeepSeek Harness Team" and updated V4 Pro with stronger agent capabilities.

V4 Pro and new pricing

The final V4 Pro launched at the same time as Harness via app, web and API. It supports the OpenAI Responses API and Codex integration, and has three reasoning modes: low, high and max. According to Forklog, DeepSeek itself has published that V4 Pro scores 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE and 74.1 on Toolathlon-Verified. The figures have not been independently verified.

Starting August 16, 2026, DeepSeek plans to introduce dynamic API pricing with peak and off-peak hours. The off-peak rate is half the peak rate.

According to PYMNTS and Bloomberg, DeepSeek has also resumed a second funding round of roughly $8 billion at a valuation of $74 billion. The first round closed in May with $7 billion at a valuation of $52 billion.