TL;DR
Jack Roberts shows how to replace Claude Code's expensive Anthropic backend with DeepSeek V4 via a proxy GitHub repo, cutting costs up to 100x while keeping the same Claude Code workflow and tool-calling functionality. The video covers setup, a live demo building a coffee website with an ROI calculator, and a recommended pipeline of Claude for design, DeepSeek for heavy lifting, and ChatGPT Codex for code review.
Key points
1
DeepSeek V4 (1.6T parameters, 1M context, MIT open weights) can replace Claude Opus 4.7 at ~$0.20/session vs ~$62.50/session using a proxy repo that intercepts Claude API calls
2
A GitHub proxy repo lets you run DeepSeek (or any OpenRouter model) inside the Claude Code harness with dual-terminal support, enabling simultaneous use of multiple models
3
DeepSeek excels at automations, scripts, algorithmic problems, unit tests, and backend work but lacks creative UI/UX taste — use Claude for design, DeepSeek for heavy lifting
4
Never paste API keys, naked URLs, or sensitive IP into DeepSeek; avoid it entirely for corporate, healthcare, or heavily regulated use cases due to data privacy concerns
5
A free alternative exists via OpenRouter by cycling through free models, though rate limiting and quality are trade-offs compared to paying a few dollars on DeepSeek directly
Actionable insights
→
Go to platform.deepseek.com, create an account, add $2-5, grab your API key, then clone the proxy GitHub repo into your IDE (e.g. Cursor/Antigravity) and configure it to route Claude Code calls to DeepSeek
→
Use a dual-terminal setup: one terminal running Claude Code for design and creative tasks, another running DeepSeek for backend, refactoring, tests, and documentation
→
For a full pipeline, use Claude for initial UI design (optionally using the awesome-design-md repo for brand design systems), DeepSeek for feature-building and code, then tag in ChatGPT Codex via your existing Plus subscription for code review
Notable quotes
“Free is not free because actually the quality is less, you get charged more.”
“If you ask this model who is it, they will basically tell you it is Claude Sonnet 4.6. That is not the actual model.”
“Deepsee by itself, if I asked it to do this, would struggle. It would probably be in crayons in the corner thinking about things.”
Worth watching?
✅
Worth watching the full video?
Watch if you want the live setup walkthrough and to see the dual-terminal workflow in action — but the key steps and decision framework for when to use each model are all captured here.
Saved you some time? The creator still deserves a like.
Want this for your own channels?
Add the channels you follow. Every new video summarised and in your inbox the moment it drops. From £4/month.
Try it free