What happened?
DeepSeek has released DeepSeek Harness, a tool aimed at turning AI models into tool-using autonomous agents, as part of a developer preview. Referred to as "dsh" for short, the tool is built on a plugin-based infrastructure called Cordis and has been shared as open source under the MIT license. DeepSeek defines an AI agent as "model + harness"; the model provides decision-making capability while the harness layer manages processes such as understanding the environment, using tools, and executing multi-step tasks.
In the tool's architecture, models, tools, skills, sessions, secure execution environments, storage systems, agent loops, scheduling, and the user interface are each treated as separate plugins. This allows developers to swap out components or add their own plugins without modifying the source code. All actions performed by the agent — system prompts, reasoning steps, tool calls, and sub-agent task assignments — are recorded in an immutable session log.
What are Harness's operating modes?
- Standard: Includes tools such as file editing, terminal, web and file search, planning, and sub-agents.
- Code: The model manages multiple tool calls together through a TypeScript program.
- Minimal: Works only with a terminal and file editor.
- Creator: Focuses on enabling developers to build their own agent configurations.
Why does it matter?
DeepSeek Harness brings an open-source alternative to a market dominated by coding-focused agent tools such as Anthropic's Claude Code and OpenAI's Codex. By enabling companies to build their own coding agents, this approach stands out as an effort to strengthen DeepSeek's position in AI-assisted software development.
Alongside the Harness announcement, DeepSeek also released DeepSeek-V4-Pro-0813, the general availability version of the V4 model family, which was previewed in April 2026. The new version is accessible via web, mobile app, and API; it natively supports the OpenAI Responses API format and offers Codex integration. According to results shared by DeepSeek, the model scored 87.9 on the Terminal Bench 2.1 benchmark, 62.7 on the DeepSWE test, and 74.1 on the Toolathlon-Verified benchmark; however, these results are based on the company's own testing and have not been independently verified.
How is pricing changing?
| Model | Input (off-peak/peak) | Output (off-peak/peak) |
|---|---|---|
| V4-Pro | $0.66 / $1.32 | $1.98 / $3.96 |
| V4-Flash | $0.22 / $0.44 | $0.66 / $1.32 |
These prices, listed per million tokens, will take effect starting at 16:00 UTC on August 16, 2026. Off-peak rates will be half of peak rates.
What's next?
DeepSeek Harness is currently in developer preview, so no timeline has been shared for a stable release of the tool. As the new pricing system comes into effect, it will become clearer to what extent developers adopt V4-Pro and V4-Flash models in agent-based workflows.