When does self-hosting beat API pricing? Here’s the math.
API costs (monthly)
| Model | Light use (1hr/day) | Heavy use (8hr/day) |
|---|---|---|
| Claude Opus | $50-150 | $400-1,500 |
| GPT-5.4 | $30-100 | $250-800 |
| DeepSeek | $3-10 | $20-80 |
| Qwen Flash | $1-3 | $5-20 |
Self-hosting costs (monthly, amortized over 2 years)
| Setup | Hardware | Monthly | Electricity |
|---|---|---|---|
| Mac Mini M4 32GB | $1,150 | $48 | $5 |
| RTX 4090 workstation | $2,500 | $104 | $15 |
| Cloud A100 (dedicated) | — | $720 | Included |
Break-even
| API spend/month | Self-host with… | Break-even |
|---|---|---|
| <$50 | Don’t self-host | Never |
| $50-100 | Mac Mini M4 | ~2 years |
| $100-300 | RTX 4090 | ~1 year |
| $300-1000 | Cloud A100 | Immediately |
| >$1000 | Dedicated server | Immediately |
The hybrid approach
Most teams should use both:
- Local models for routine work (free)
- Cheap APIs for medium tasks (DeepSeek at $0.27/1M)
- Premium APIs for hard problems (Claude Opus)
This is the approach we use in our AI Startup Race — cheap models for routine sessions, premium for complex tasks.
See our cheapest AI coding setup and cost reduction guide for detailed strategies.
Related: LLM Cost Calculator Guide · How to Reduce LLM API Costs · Self-Hosted AI vs API · Serverless vs Dedicated GPU