DeepSeek Ships V4 Pro, Open-Sources Harness to Rival Claude Code

DeepSeek has taken its flagship model out of preview and paired it with an open-source coding-agent runtime, marking its most direct challenge yet to Anthropic's Claude Code. The Chinese lab rolled out DeepSeek-V4-Pro-0813 in general availability across its app, web, and API, while simultaneously releasing DeepSeek Harness v0.1 under the MIT license on GitHub.

What happened

DeepSeek-V4-Pro now carries significantly stronger agent capabilities than its preview. On DeepSWE, which measures real-world software-engineering problem solving, the model jumped to 62.7 from the preview's 12.8. It also posts 87.9 on Terminal Bench 2.1, 61.5 on NL2Repo, and 83.3 on Cybergym, positioning it as a serious coding-agent contender. The API now natively supports the OpenAI Responses API format and is specifically adapted for Codex, with a one-click configuration script, according to DeepSeek's changelog.

Alongside the model, DeepSeek Harness v0.1 is an open-source agent harness built in TypeScript on a vendored fork of Cordis, with an "everything is a plugin" architecture. The terminal command dsh lets developers run multi-step tasks across files, commands, and the web. The repository amassed more than 24,000 GitHub stars within a day.

Why it matters

Harness is the more strategically important of the two releases. It moves DeepSeek from a model provider into the developer-tooling layer, giving builders a free, self-hostable alternative to Anthropic's Claude Code and the newer Claude Cowork. Combined with V4 Pro's low API pricing, it makes an end-to-end open agent stack viable for teams that want to avoid vendor lock-in.

DeepSeek also introduced three thinking-effort levels (low, high, and max) across V4 Pro and V4 Flash, and is moving to peak/off-peak pricing effective August 16, with off-peak output tokens priced at half the peak rate.

Why it matters for builders

For n8n Lab readers, DeepSeek Harness is worth watching because it competes on the exact layer where agent builders live: the harness that turns a model into a tool-using coder. An MIT-licensed, plugin-based runtime under active development signals that coding-agent infrastructure is becoming commoditized, which pressures incumbents on price and forces them to differentiate on orchestration and reliability rather than raw model quality.