DeepSeek Harness and updated V4-Pro put coding agents and pricing in focus
DeepSeek released its modular open-source Harness preview and updated V4-Pro, while raising API rates from August 16.
By Renata Fuchs · Policy Reporter
· 3 min read
DeepSeek Harness V4-Pro is now DeepSeek’s combined pitch to developers: an MIT-licensed agent-runtime preview alongside an updated official V4-Pro API release. The company also plans to replace flat API pricing with peak and off-peak rates on August 16 at 16:00 UTC, raising listed prices for several common agent-workload token categories.
Harness v0.1 moves DeepSeek beyond supplying a model endpoint and into the software layer that gives an AI agent access to tools, files, sessions and execution loops. V4-Pro-0813, meanwhile, is an updated official version of a model family DeepSeek introduced in preview in April, rather than V4-Pro’s first release.
Is DeepSeek Harness a replacement for Claude Code or Codex?
Not on the evidence available today. Harness covers the core coding-agent loop: it can inspect repositories, edit files, run shell commands, search files and the web, keep plans, call skills, delegate work to subagents and apply approval policies. But DeepSeek describes it as a developer preview and warns that compatibility-breaking changes are expected.
That makes the product an open-source alternative at the agent-runtime layer, rather than a demonstrated replacement for the broader products sold by Anthropic and OpenAI. The available documentation does not describe a DeepSeek-managed hosted background-agent service or a finished GitHub-native pull-request workflow, both areas where Claude Code and Codex have more established offerings.
DeepSeek built Harness on Cordis, a plugin framework whose premise is that major parts of an agent can be swapped out. Models, tools, skills, sessions, sandboxes, filesystems, orchestration and the user interface are designed as replaceable components. The company says developers can run it locally through a web interface, use a headless command option or work through a Python SDK. It can be launched through npm or built from source.
That architecture gives teams more latitude to choose models and modify the runtime than a vertically integrated coding product. It also puts more responsibility on those teams to test reliability, security controls and integration behavior before using a preview tool in production.
V4-Pro-0813 is available through DeepSeek’s web interface, mobile app and API. According to DeepSeek, it adds native support for the OpenAI Responses API and Codex integration; the company also offers low, high and max reasoning-effort settings. DeepSeek has published improved agent-benchmark results for the release, but those vendor results are not a substitute for testing on representative engineering work. Teams comparing models should use a structured AI model evaluation that measures their own repositories, tools and failure tolerance.
How do DeepSeek’s new V4-Pro API prices work?
The new schedule splits usage into peak and off-peak periods. For V4-Pro, off-peak cache-miss input pricing rises from $0.435 to $0.66 per million tokens, while output rises from $0.87 to $1.98. Peak input and output prices will be $1.32 and $3.96 per million tokens, respectively.
Cache-hit pricing changes more sharply: from $0.003625 to $0.022 per million tokens off-peak and $0.044 at peak. That matters for coding agents that repeatedly retrieve the same code or other context. A team’s actual bill will still depend on its mix of cached input, new input, output tokens and the time its requests run.
This story draws on original reporting from VentureBeat.
