If you're paying for ChatGPT Plus, Claude Pro, or any AI subscription, you've probably noticed something frustrating: your bill keeps growing even though you're on a "fixed" monthly plan.
Here's why.
The Token Problem
If you're paying for ChatGPT Plus, Claude Pro, or any AI subscription, you've probably noticed something frustrating: your bill keeps growing even though you're on a "fixed" monthly plan.
Here's why.
Tokens are the basic unit AI models use to process text. When you send a prompt, your text gets broken into tokens. When AI responds, its output gets counted in tokens too. Every. Single. Token. costs money.
For context: 1 token ≈ 0.75 words. A typical conversation with Claude might use 3,000-5,000 tokens. Do that daily, and you're looking at real money.
The real problem? Those "unlimited" plans aren't unlimited. They have hidden ceilings, and once you hit them, the overage charges hit hard.
Why Your Subscription Isn't the Answer
You might think an annual or monthly subscription protects you. Here's what actually happens:
- You sign up for $20/month
- A few months in, you enable "extra usage" because you're close to your limit
- That budget feature looks great on paper
- Next thing you know, you're staring at a $100 charge
The temptation is real. "I'll just use a bit more" turns into "why is my card charged again?"
The bigger issue: you're using a sledgehammer to crack a nut. Using Claude Opus 4.6 to write a simple function? That's like hiring a Michelin-star chef to make a PB&J.
My Local Setup - Free Models That Actually Work
I got tired of the token treadmill, so I built out local models using Ollama. Here's what runs on my machine:
| Model | Size | What I Use It For |
|---|---|---|
| DeepSeek 6.7B | 3.8 GB | Code generation, quick fixes, refactoring |
| Qwen 2.5 Coder 14B | 9.0 GB | Blog writing, explanations, documentation |
| StarCoder 2 3B | 1.7 GB | Fast, lightweight tasks, simple scripts |
All free. All local. No token limits.
The key insight: these models handle 80% of my work just fine.
The Hybrid Strategy - When to Pay, When Not To
Here's my framework:
Use free local models for:
- Writing simple functions
- Quick code reviews
- Documentation updates
- Explaining code patterns
- Routine refactoring
Pay for premium models when:
- Complex architecture reviews that need deep reasoning
- Truly technical debugging across multiple systems
- Tasks requiring the absolute best reasoning capabilities
- When you need the latest model with the largest context window
The rule is simple: use the right tool for the right job.
Don't use Claude Opus to write a for-loop. Don't use GPT-4 to explain what a function does. Local models are more than capable for everyday tasks - save the expensive models for when you genuinely need them.
Conclusion
AI subscriptions are tempting because they promise simplicity. But understanding your actual usage - and choosing the right model for each task - saves real money.
Local models have come a long way. Qwen, DeepSeek, and StarCoder are genuinely good enough for most development work. The only time I fire up Claude now is for genuinely complex architectural decisions.
Use judgment. Use the right tool. Don't let tokens eat your budget.
Questions? Reach out or drop a comment below.