DeepClaude Open-Source Script Emerges: Reducing Claude Code Execution Costs by 17x Using DeepSeek V4 Pro

On May 4, 2026, a new open-source script named 'DeepClaude' was released, shaking up the economic landscape of AI-powered software development. This script claims to reduce the costs of running Anthropic’s high-performance development agent, 'Claude Code,' to approximately 1/17th of its original price. Developers can now leverage low-cost, high-efficiency models like DeepSeek V4 Pro as a backend while maintaining Claude Code's sophisticated agent loop.

The emergence of this script is significant as it acts as a bridge between Anthropic’s closed ecosystem and affordable large language models (LLMs) from China. DeepClaude supports various alternative backends, including OpenRouter and Fireworks AI, in addition to DeepSeek V4 Pro. This provides developers with the flexibility to choose inference models based on their budget and performance needs, which is expected to be a strong incentive, particularly for large-scale projects requiring frequent API calls.

DeepClaude replaces Claude Code's expensive Anthropic backend with DeepSeek V4 Pro, OpenRouter, or Fireworks AI, drastically reducing billing costs while maintaining the agent loop.

Technically, DeepClaude operates through a proxy module and a CLI (Command Line Interface) wrapper. By intercepting Anthropic’s proprietary API calls and redirecting them to DeepSeek or OpenRouter APIs, users can continue using the Claude Code interface while the actual computation is performed on much cheaper models. This approach is evaluated as a strategy to maximize cost efficiency without disrupting existing workflows.

Technical Architecture: Model Swapping via Proxy Modules

According to the technical specifications of DeepClaude published on GitHub, the development process of this system includes an interesting case of collaboration. Commit records from May 3, 2026, confirm that the 'Claude Opus 4.6' model participated as a co-author during the writing of the proxy code. The fact that a high-performance AI model contributed to creating a tool to lower its own execution costs has become a major topic of discussion within the tech community.

• proxy/start-proxy.js: Serves as the core CLI wrapper for executing the proxy module
• Remote control screenshots including the DeepSeek V4 Pro execution screen and related documentation updates completed
• 17x cost reduction effect formalized through the initial release on May 3, 2026

In terms of inference economics, the price difference between Anthropic and DeepSeek is stark. As of March 2026, the price per input token for Anthropic’s flagship model, Claude Opus, is around $5 per 1 million tokens, whereas DeepSeek V3.2, released in April 2026, costs only $0.14. DeepClaude strategically exploits this price gap to achieve over 95% savings in actual computation costs while maintaining the Claude Code user experience.

Analyses suggesting that low-cost models are not necessarily inferior in performance are gaining traction. According to SWE-Bench results, a software engineering benchmark, China’s Kimi K2.5 model recorded a success rate of 76.8%, surpassing Claude Sonnet’s approximately 72%. This suggests a practical possibility that developers can save costs while potentially gaining in code quality through DeepClaude.

Market Implications and Geopolitical AI Competition

The proliferation of this script also illustrates the discrepancy between the U.S. government’s assessment of Chinese AI and the reality on the ground in development. While CAISI, under the U.S. National Institute of Standards and Technology (NIST), recently evaluated DeepSeek V4 Pro as lagging behind U.S. models, some critics argue that the benchmark results are biased in favor of specific U.S. models. Actual developers are reacting more sensitively to the practical efficiency and cost savings achievable through tools like DeepClaude rather than political judgments.

DeepClaude users can still utilize the latest features recently introduced to Claude Code by Anthropic. According to updates from late April and early May 2026, Claude Code supports a new /usage command that integrates the previous /cost and /stats commands, and user settings are permanently stored in the ~/.claude/settings.json file. Even in the DeepClaude proxy environment, these settings are correctly applied according to project-specific policy priorities, maintaining user convenience.

However, the long-term survival of DeepClaude depends on Anthropic’s future response. If Anthropic blocks API access via third-party proxies based on its terms of service or more tightly couples Claude Code’s agent logic with its own backend, 'model swapping' methods like DeepClaude could face restrictions. Nevertheless, the attempt to find cost-effective alternatives within the developer community is expected to continue, likely accelerating price competition in the AI model market.