What happened?
On August 13, 2026, DeepSeek introduced the official release of its flagship model DeepSeek-V4-Pro along with a new open-source agent framework, DeepSeek Harness v0.1. Harness is available as a developer preview download on GitHub, with the source code shared by the company.
DeepSeek states that Harness is built on the Cordis framework, designed around the principle that "everything is a plugin." Components such as models, tools, skills, sessions, sandboxes, file systems, and interfaces are all offered as replaceable plugins. In Standard mode, the tool supports functions like file editing, shell access, file and web search, planning, delegating tasks to sub-agents, and approval policies.
Why does it matter?
DeepSeek is no longer competing solely on model intelligence and token prices; it is now entering the agent infrastructure layer dominated by products like Anthropic's Claude Code and OpenAI's Codex. This layer determines how models use tools, modify files, and carry out long-running agent workflows.
Harness is not yet presented as a stable production platform; the GitHub repository carries a warning that "there will be breaking changes." Even so, the repository gathered around 27,500 stars and 2,000 forks on release day, signaling developer interest — though these fast-moving numbers should be viewed as a snapshot rather than a true adoption metric.
On the V4-Pro side, DeepSeek says it has significantly improved agent capabilities and now natively supports OpenAI's Responses API. The model is optimized for one-click setup with OpenAI's open-source Codex tool. Developers can access the latest version via the API by calling deepseek-v4-pro without changing the model ID.
Comparison: Harness, Claude Code, and Codex
| Feature | DeepSeek Harness | Claude Code | OpenAI Codex |
|---|---|---|---|
| Repo reading/editing/testing | Supported | Supported | Supported |
| Permissions and sandbox control | Configurable via plugins | Mature, built-in system | Detailed sandbox and approval controls |
| Hosted background agents | Undocumented | Supported | Supported |
| Model selection | DeepSeek, Anthropic, OpenAI, and compatible endpoints | Primarily Claude | Primarily OpenAI models |
| Product maturity | Developer preview | Established commercial product | Established commercial product and open-source CLI |
| License | MIT | Commercial product | Codex CLI open source; cloud services are a managed product |
What's next?
Starting August 16, 2026 at 16:00 UTC (19:00 Turkey time), DeepSeek will remove its current flat API pricing and switch to peak and off-peak hour pricing. According to the company, even the discounted off-peak cache-miss and output prices will be notably higher than today's rates.
Since Harness is still in the developer preview stage, breaking changes are expected to continue. DeepSeek offers options to run the tool via npm with npx @deepseek-ai/dsh web or by building it from source.
What we know
- The V4 family was first introduced as a preview in April 2026; the August 13, 2026 release represents the official transition.
- V4-Flash has 284 billion parameters, activating 13 billion parameters per token.
- Harness offers three reasoning levels: Non-think, Think High, and Think Max.