DeepSeek Releases V4 Pro and Open-Source Harness

DeepSeek has released the production version of its V4 Pro model alongside DeepSeek Harness, an open-source agent runtime designed to handle the tools, sessions, sandboxes and execution loops around the model. 

The V4 Pro is now available through DeepSeek’s API, web and app, while Harness has been released as a Developer Preview under the MIT licence.

V4 Pro is a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters, according to DeepSeek’s earlier technical documentation.

DeepSeek first released V4 Pro in April as a preview, alongside V4 Flash.

It retains a 1-million-token context window and is positioned around coding, tool use, and agentic workloads. Independent testing by Artificial Analysis gave the production model an Intelligence Index score of 53, compared with 40 for V4 Flash.

The production release also comes with a major change to DeepSeek’s API pricing. From August 16, the company will introduce peak and off-peak rates for V4 Pro and V4 Flash, with prices varying by model, token type and time of use. 

Peak-hour rates for V4 Pro will be more than four times its current output-token price, while the overall increases across the two models range from 50% to 1,100%, according to DeepSeek.

DeepSeek is also moving beyond the model itself with Harness, which provides the execution layer for agentic applications. 

The Developer Preview is built around Cordis, a plugin system in which models, tools, sessions, sandboxes, storage, agent loops, and user interfaces can be replaced or extended as plugins. 

The system includes four operating modes—Standard, PTC (Programmatic Tool Calling), Minimal, and Creative. Standard provides a broader toolset for general agent tasks, while Minimal restricts the environment primarily to shell and file-editing capabilities for model evaluation. 

PTC allows the model to generate code that orchestrates multiple tool calls, while Creative is designed for experimentation with the runtime and Cordis plugins.

Harness also uses an append-only session log as the source of truth for an agent run. The system can reconstruct conversations and execution state from recorded events, allowing runs to be resumed, branched, and replayed. Tool calls, results and other execution events can be inspected through its trajectory interface.

The release puts DeepSeek into the same infrastructure layer as agent frameworks surrounding models from OpenAI and Anthropic. 

Rather than treating an agent as only a model with tool access, Harness packages the control loop, context management, state persistence and tool execution into a modifiable runtime. 

DeepSeek’s own API documentation had already positioned V4 as compatible with external agent frameworks including Claude Code, OpenClaw, and OpenCode.