DeepSeek Harness: free coding agent and pricier V4-Pro

What's new

DeepSeek released Harness v0.1 on August 13, 2026. It is an open source project on GitHub under the MIT license, so anyone can freely download, use and modify it. Harness is not an AI model but an agent runtime: a tool that lets a language model work as an autonomous coding assistant. It can read files, run commands and solve multistep tasks.

A new version of the model arrived at the same time: V4-Pro-0813. It has a context window of 1 million tokens (chunks of text), meaning room for a lot of material at once. It supports the OpenAI Responses API with Codex integration and has three reasoning levels: low, high and max.

Harness is a developer preview and not intended for production. Breaking changes are expected, and the session format has no stability guarantee.

What's clever

Harness is built on the plugin framework Cordis. Models, tools, sessions, sandboxes and the user interface are all plugins that can be swapped out. It is optimized for DeepSeek's V4-Pro and V4-Flash, but according to the coverage it can use other language models through adapter plugins.

It is easy to try: the command npx @deepseek-ai/dsh web opens a web interface on your own computer.

Interest has been high. According to Pasquale Pillitteri's coverage, the repository reached 27,000 stars on GitHub within a few hours, and according to The Decoder, 712 projects signed up for the beta within three days. The team was formed in May 2026 and is led by Cui Tianyi, who came from Jane Street in March 2026.

Cheaper or better?

Better, but not cheaper. According to The Decoder, V4-Pro's score on Terminal Bench 2.1 rose from 72.1 to 87.9, and on DeepSWE from 12.8 to 62.7. The figures are probably DeepSeek's own. On the Artificial Analysis Intelligence Index, V4-Pro scores 53 points, level with GLM-5.2, while Claude Opus 5 sits at 63.

Prices have gone up and now depend on the time of day. Peak time is Monday to Friday 01-04 and 06-10 UTC, except Chinese public holidays. Per million tokens, it now costs:

- Input without cache hit: $0.66 off-peak, $1.32 at peak (previously 0.435) - Output: $1.98 off-peak, $3.96 at peak (previously 0.87) - Input with cache hit: $0.022 off-peak, $0.044 at peak (previously 0.003625)

A cache hit is when the model reuses text it has seen before. Here the percentage increase is the largest. The new prices apply from August 16, 2026 at 16:00 UTC.

Compared with Claude Code, the difference is that Harness is open source with swappable parts, while Claude Code is proprietary and tightly tied to Anthropic's models. Independent tests of the two do not appear in the sources. A price for Harness itself is not relevant, since the program is free, but you pay for the model you connect to it.

What it's good for

Harness is aimed primarily at developers and technical teams who want to experiment with an autonomous coding agent without being locked in to a single vendor. Because it is a preview, it is not yet suited for critical systems.

V4-Pro can deliver up to 384,000 tokens of output and supports JSON output and tool calls, but not images. Concurrent calls are limited to 500, versus 2,500 for Flash. Companies with large, heavy workloads can also benefit from scheduling them outside peak time to get the low price.

Harness is freely available on GitHub, and the API can be used through DeepSeek. The sources do not say whether special terms apply in Denmark or the EU.