DeepSeek challenges Anthropic: open Harness and final V4 Pro with new pricing plans
The Chinese laboratory DeepSeek is once again reshaping the landscape of the AI development market by unveiling two significant products at once. These are the open environment for creating AI agents, Harness v0.1, and the final version of the flagship model V4 Pro. This is not just an incremental update, but a strategic move aimed at directly competing with Anthropic's ecosystem and their popular tool, Claude Code.
Harness: A New Philosophy of Agentic Development
The key innovation is the Harness framework, distributed under the MIT license. Unlike traditional approaches where the model merely generates code fragments, Harness acts as a full-fledged "thin layer" on top of the LLM, endowing the agent with the ability to independently interact with the codebase, terminal, and external services. This architectural solution transforms the model from a passive generator into an active executor capable of building long and complex chains of actions.
It is important to emphasize that Harness is positioned not as a narrowly specialized tool, but as a minimalist and flexible framework. Developers can connect their own models to it and adapt it to specific workflows. Based on my information, this very environment was used for internal testing of V4 Pro's agentic capabilities on public benchmarks, which indicates the company's high level of trust in its own framework.
V4 Pro: A Bet on Agentic Capabilities and Compatibility
The final release of DeepSeek V4 Pro marks a significant leap in performance. The model has received native support for the OpenAI Responses API and integration with Codex, making it an almost seamless replacement for teams already using OpenAI's infrastructure. A flexible reasoning system with three modes (low, high, max) allows developers to balance between speed and depth of analysis depending on the task.
The published test results look impressive: 87.9 points in Terminal Bench 2.1, 62.7 in DeepSWE, and 74.1 in Toolathlon-Verified. These figures demonstrate that DeepSeek is not just catching up, but in some aspects surpassing Western competitors in the field of agentic AI, especially given the openness of the source code.
Changes in Pricing Policy
Along with the release, DeepSeek announced a transition to a dynamic pricing model with peak and off-peak hours. In effect, this means a twofold increase in cost during hours of highest load, which will take effect from August 16. Despite this, the price remains extremely attractive compared to OpenAI and Anthropic offerings, maintaining DeepSeek's status as the most affordable provider of high-performance models.
My analysis: The release of Harness is a much more subtle move than it seems at first glance. DeepSeek is not just creating an analog of Claude Code, but is building an entire ecosystem around V4 Pro. The open code and compatibility with the OpenAI API are a powerful catalyst for the influx of developers seeking independence from a specific vendor. In the long term, it is the combination of "model + open agentic environment" that could become DeepSeek's main trump card in the fight for developers' minds and wallets, especially against the backdrop of growing dissatisfaction with the policies of closedness and pricing in Big Tech.