GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

DeepSeek Harness v0.1: MIT-Licensed Coding Agent With Swappable Plugins Enters the Claude Code Fight

DeepSeek shipped Harness v0.1 under the MIT license on August 13. It is a developer preview of the lab’s own coding agent — a direct competitor to Claude Code and OpenAI Codex — built on a plugin architecture called Cordis that makes every feature a swappable module.

The team behind it moved fast. Cui Tianyi, who joined DeepSeek in March 2026 after a stint at Jane Street, leads the Harness project. The lab publicly announced the team’s formation and began shipping within the same year.

Architecture

Harness is built around the Cordis plugin system. The design choice is deliberate: rather than a monolithic agent loop like Claude Code’s while-loop architecture, every component in Harness is a plugin. Context management, tool routing, memory, and execution are all independently replaceable. The system ships with native support for the OpenAI Responses API and integration with Codex, which means it can be pointed at multiple model backends without modification.

The framework is designed for developers who want to build their own agent workflows on top of the core, rather than accepting a black-box agent.

The Concurrent Price Hike

The open-source move comes alongside a significant API pricing change. Cache hit costs are jumping to six times their current rate — the largest component of the transition that hits agent workflows hard.

Agent workflows repeatedly read the same files, context, and tool outputs. Cache hits are where most of the cost savings in agentic applications come from. A 6x increase in cache pricing is effectively a targeted tax on exactly the use case Harness is designed to enable. For teams running DeepSeek V4-Pro as the backend model in agent pipelines, the economics of the transition are more complex than the headline open-source announcement suggests.

Context

V4-Pro moved from preview to general availability on the same day. The 1.6T MoE model — which was already the subject of a separate release article — is now the default backend for Harness v0.1.

On the Artificial Analysis Intelligence Index, V4-Pro sits at intelligence score 50, in the same tier as several mid-frontier models. Its cost advantage — $0.87/M output after the 75% price cut locked in earlier — was the primary value proposition. The cache hit price hike complicates that math for agent-heavy customers, even as Harness makes the developer experience more accessible.

Why It Matters

The coding agent market has two dominant players: Claude Code, which controls the startup segment, and Codex, which has enterprise momentum following the ChatGPT Work integration. Both are proprietary. Harness is the first serious open-source competitor from a frontier lab.

The MIT license means any developer or company can run, fork, and modify Harness without restriction. That’s a different proposition than building on Claude Code’s API or Codex’s managed environment. Whether Harness can match those tools on real-world task performance is the remaining unknown — DeepSeek has not released Harness-specific benchmark numbers against SWE-Bench or comparable agentic evals.