DeepSeek challenges the giants: open-source Harness framework announced and final V4 Pro release
The Chinese laboratory DeepSeek is once again drawing market attention by unveiling two significant products at once. This refers to the launch of Harness v0.1 — an open-source development environment for AI agents — as well as the release of the final version of its flagship model, V4 Pro. This move is clearly aimed at strengthening its position in the competitive battle against industry giants such as OpenAI and Anthropic.
The key new release is Harness, which is positioned as an open alternative to Anthropic's popular tool, Claude Code. The framework is distributed under the MIT license and is already available to developers in a preliminary version on GitHub. At its core, Harness is a thin layer on top of large language models (LLMs) that provides the agent with a full arsenal of tools for working with codebases, the terminal, and external services. It is not just a code generator, but a full-fledged environment for autonomously executing complex sequences of actions.
Alongside the framework, DeepSeek also presented the final version of its V4 Pro model. The developers significantly enhanced its agentic capabilities and added the long-awaited support for the OpenAI Responses API. The model is now available in the web version, the app, and via the API, offering users flexible configuration of reasoning levels: low, high, and max.
Agentic Power and New Benchmarks
DeepSeek is betting on the synergy of "model + agentic shell," and it was precisely Harness that was used to test the agentic capabilities of V4 Pro. The results are impressive: the model scored 87.9 points in Terminal Bench 2.1, 62.7 in DeepSWE, and 74.1 in Toolathlon-Verified. These figures demonstrate serious progress in the model's ability not only to reason but also to effectively interact with external tools while executing long chains of operations.
Changes in Pricing Policy
With the full release of V4 Pro, DeepSeek is also revising its pricing policy. The company is introducing a dynamic pricing model that includes peak and off-peak hours. During off-peak periods, the cost will be half as much — a clear step aimed at attracting developers and optimizing infrastructure load. The new rates take effect on August 16.
My expert perspective: The release of Harness is not just another open-source project. It is a strategic move that allows DeepSeek to build its own ecosystem, independent of competitors' developments. The combination of the powerful V4 Pro model with a flexible and open framework could become a serious argument for developers seeking alternatives to expensive proprietary solutions. The AI agent market is entering a phase of intense competition, and DeepSeek has every chance to carve out its niche here.