'deepclaude,' a Claude Code agent loop using DeepSeek V4 Pro, can reduce costs to a mere 1/17th of the original.
' deepclaude ' is a tool that allows you to switch the model used in the background to DeepSeek V4 Pro or similar while maintaining the ease of use of Claude Code's autonomous coding agent. The key point is that it can reduce costs by up to 1/17th compared to using Anthropic models, while maintaining the same user experience as Claude Code.
GitHub - aattaran/deepclaude: Use Claude Code's autonomous agent loop with DeepSeek V4 Pro, OpenRouter, or any Anthropic-compatible backend. Same UX, 17x cheaper. · GitHub
Claude Code is a coding agent that offers features such as reading and editing files, executing bash, Git operations, and autonomous multi-step work loops. However, using Claude Code requires a capped plan of $200 per month (approximately 32,000 yen).
deepclaude is a mechanism that replaces the model calling destination with DeepSeek V4 Pro and other similar services while maintaining the tool loop and user experience of the Claude Code CLI. By sending API calls from Claude Code to backends such as DeepSeek, OpenRouter, Fireworks AI, and Anthropic, it allows you to use different models with a similar user experience.
The features that can actually be used include reading, writing, and editing files, executing Bash/PowerShell, searching with Glob and Grep, running multi-step autonomous loops, launching subagents, Git operations, and project initialization. In other words, deepclaude is not simply about changing the chat destination to DeepSeek, but rather a tool that aims to run the development work loops of Claude Code in a less expensive model.
However, deepclaude does not completely replace Claude Code itself. DeepSeek's Anthropic-compatible endpoint does not support image input, and the MCP server tool is not supported via the compatibility layer. Also, prompt caching by Anthropic's cache_control is ignored, and DeepSeek's automatic caching is used instead.
Furthermore, while DeepSeek V4 Pro performs similarly to Claude Opus in about 80% of routine tasks, Claude Opus is considered superior in the remaining 20% of tasks requiring complex inference. Therefore, it is envisioned that users will use DeepSeek for normal tasks and switch back to Anthropic for more difficult problems.
Furthermore, deepclaude also provides a feature to switch backends during a session. A local proxy runs on localhost:3200, routing model API calls from Claude Code to the currently selected backend, allowing you to switch between DeepSeek, OpenRouter, Anthropic, etc., without restarting.
As of the time of writing, DeepSeek V4 Pro costs $0.44 (approximately 69 yen) per million input tokens (cache misses) and $0.87 (approximately 136 yen) per million output tokens. In contrast, Anthropic's Claude Opus 4.7 costs $5 (approximately 780 yen) per million input tokens and $25 (approximately 3920 yen) per million output tokens. The price of DeepSeek V4 Pro is a 75% discount until May 31, 2026, but even after the discount ends, DeepSeek V4 Pro will still be the lower-cost option. deepclaude can be seen as an experimental option for developers who want to reduce model usage fees while maintaining the Claude Code development experience.
Related Posts:
