Switching from GitHub Copilot to OpenCode (DeepSeek V4 Pro)
DeepSeek V4 Pro (Max) has a very low cost per task and intelligence close to Sonnet 4.6 (Max). Source: Artificial Analysis
I used GitHub Copilot for 3 months. In the last month, I ran out of credits halfway through. I upgraded from Copilot Pro+ to Copilot Max at $100/month, then used up half of those credits in a single week. It was too expensive, so I switched to OpenCode.
OpenCode
OpenCode’s Go plan costs $5 for the first month and $10/month after. It also includes other cheap models besides DeepSeek. Setup took under a minute:
- Sign up at Opencode and subscribe to the Go plan for $5
- Go to the Opencode Dashboard and copy your API key
- Install the OpenCode for Copilot Chat extension in VS Code
- Press
Ctrl+Shift+P, search OpenCode Go: Set API Key, and paste your API key - Open GitHub Copilot Chat (
Ctrl+Alt+I) - Choose DeepSeek V4 Pro
- Set Thinking Effort to Max
Performance
DeepSeek V4 Pro (Max) feels as good as Claude Sonnet 4.5. Before this, I would switch models for each task:
- Quick questions and brainstorming: GPT-5.4 Mini and MAI-Code-1-Flash
- Everyday coding: Gemini 3.5 Flash and GPT-5.3-Codex
- Persistent bugs and complex tasks: Claude Opus 4.8, which works well but uses credits fast
Now I use DeepSeek V4 Pro (Max) for all three. Set it and forget it. DeepSeek V4 Flash is even cheaper and fine for chatting, but Pro is already so cheap I don’t bother switching.
Cost
For 1/10 the price, I get the same work done. I was paying $100/month for Copilot Max and still running out of credits. Now I pay $10/month, with credits to spare.
Extras
- Short, clean responses: DeepSeek keeps replies short. After writing code, it says things like “Fixed: updated config” or “Done. added endpoint.” Copilot models tend to use wordier, agent-style language like “I’ve done this…” or “Let me handle that…”
- Always builds: DeepSeek builds the project every time without me asking. Copilot models were hit or miss. Sometimes they build, sometimes not.
CAUTIONDeepSeek is a Chinese model. It can be useful and cost-effective for personal projects, small apps, or demos at a hackathon. However, if you work at a US company, especially one with government contracts, you may need to avoid it for compliance reasons. Always use it at your own discretion.