I've wasted a lot of money on AI subscriptions.

When Claude Code first came out, it was impressive—but $200 a month still stung when the bill arrived. So I did the math: do I really need the most expensive model for every task?

Turns out, no.

My current setup costs around 400 RMB per month. That's way cheaper than a single Max subscription. Here's what I use, in case it helps anyone else in the same boat.


My Model Mix Strategy

The core idea is simple: easy tasks get the cheap model, hard tasks get the premium one.

MiniMax (Daily Workhorse)

For routine stuff—summarizing docs, filtering emails—I don't need the strongest model. I just need it to be available and affordable.

I use MiniMax's token plan, 98 RMB/month. The M2.7 high speed tier is plenty fast enough.

What it handles well:

  • Document and email summaries ✅
  • Small development tasks (UI tweaks, utility scripts) ✅
  • Complex system design, multi-module integration ❌

Anything beyond a "big module" level implementation and it starts falling short—context window and reasoning just don't cut it. When that happens, I switch to the next tool.

I also have MiniMax wired into Claude Code to run various skills like humanizer. Claude Code isn't doing dev work here—it's just a skill execution engine.

MiniMax token plan usage screenshot

GitHub Copilot Pro+ (Primary Code Development)

For coding, it's Copilot all the way.

Copilot Pro+, $39/month, gives me 1500 user message credits per month (based on my messages only—AI replies and intermediate tool calls don't count). I rarely run out.

I use GPT-5.4 (xhigh)—xhigh is the reasoning depth level.

Two Copilot skills I rely on:

  • UI/UX Pro Max — designs frontend interfaces and interaction logic. I let it generate UI proposals instead of grinding them out myself.
  • Test Driven Development — writes test cases before any change, then runs them automatically. Since installing this skill, my development rework rate dropped by over 90%. Features basically don't need secondary revisions; occasionally just need style tweaks or clarification on my part.

Copilot settings screenshot


OpenClaw (Personal AI Assistant)

OpenClaw is my most frequently used AI interface day-to-day, running on MiniMax, accessed through Discord.

Discord channels are organized by function—each group handles its own thing: email summaries, news feeds, site management, note-taking.

Daily Email Digest

Every morning, OpenClaw automatically reads my emails from the previous day and filters out what actually needs attention. Not just a list of subject lines—it extracts action items and key info from each email.

9 AM News Push

AI news and international finance, pushed at 9 AM. Same channel, organized by category.

Personal Site Management

Integrated with Google Search Console access. OpenClaw analyzes click-through rates and helps me continuously optimize SEO. I can ask it anytime what it's done and how it's performing.

Second Brain

Note-taking, Managing positions, managing tasks—all through OpenClaw. I've built software that lets OpenClaw handle these operations inside it.

Discord OpenClaw channel screenshot


Cost Comparison

Plan Monthly Cost
Claude Code Max $200 ≈ 1450 RMB
Copilot Pro+ $39 ≈ 280 RMB
MiniMax 98 RMB
My Mix (MiniMax + Copilot Pro+) ~378 RMB ≈ $52

MiniMax covers most of what I do. Copilot Pro+ only kicks in for code. My actual monthly spend is just those two subscriptions combined—378 RMB.


Summary

Expensive AI tools aren't always better. The key is matching the right task to the right model.

Simple summaries go to MiniMax. Code development goes to Copilot. Personal agent tasks go to OpenClaw. This mix has been stable, affordable, and gets the job done.

If you're thinking about optimizing your AI costs, hope this gives you something to work with.