Chapter 10: How to Stay Within Free Tiers

With great deployment comes great responsibility, Chapter 10 is all about smart cost control and usage monitoring — the key to sustainably running your AI apps, especially when using paid APIs (like OpenAI/Replicate) or free-tier compute (Railway/Hugging Face/Vercel). Let’s protect your wallet while scaling your impact.

---

10.1 Why Cost Optimization Matters

Even small AI/ML apps can rack up unexpected costs due to:
- Token overuse (OpenAI, Claude)
- Repeated image inference (Replicate, Stability)
- Exceeding platform quotas (Railway, Vercel)

Goal: Ensure your app remains free or low-cost until you're ready to scale.

---

10.2 Free Tier Comparison Recap (Limits)

Platform	Monthly Free Tier	Key Limitations
OpenAI	Free trial ($5–$18, one-time)	After trial, pay-per-token
Replicate	$10 in credits (one-time)	Pay-per-inference after
Hugging Face	Free Spaces + CPU only	No GPU, low RAM unless PRO
Railway	500 hrs/month, 1GB deploy	Cold starts, no GPU
Vercel	100GB bandwidth, unlimited deploys	Bandwidth limit for image-heavy apps

You can deploy multiple apps under one free plan — just rotate projects if needed!

---

10.3 Control API Costs with Smart Code

1. Set Prompt Length Limits (for OpenAI)

       if len(prompt) > 250:
           return {"error": "Prompt too long"}
   

2. Add a Global Cooldown (e.g. 10 seconds)

    import time
    last_used = 0
    def safe_generate(prompt):
        global last_used
        now = time.time()
        if now - last_used < 10:
            return {"error": "Please wait before generating again."}
        last_used = now
        # call your OpenAI API

1. Use Small Models First

Model	Est. Cost per 1K tokens
gpt-3.5-turbo	~$0.0015
gpt-4	~$0.03–$0.06
openai/text-davinci-003	~$0.02

Use gpt-3.5-turbo by default for text.

---

10.4 Optimize Image Inference (Replicate)

Strategy	Result
Cache output URLs	Save storage and bandwidth
Avoid large image inputs	Resize before sending
Use “Preview” mode in demos	Lower-res output = cheaper
Bundle image post-processing	Avoid second inference step

You can also pre-generate results for demos to avoid live inference costs.

---

10.5 Monitoring Usage & Logs

Platform	Tool / Page	What to Check
Hugging Face	Spaces → Logs	Inference errors, load time
Railway	Project Logs	Backend errors, API calls, cold starts
OpenAI	Usage Dashboard	Token usage and cost breakdown
Replicate	Billing Page + History	# of model runs, average cost
Vercel	Analytics (Pro only)	Bandwidth, requests

Check logs after every major feature update.

---

10.6 Add Logging to Your Backend

utils/logger.py

    import datetime
    def log_usage(endpoint: str, input_data: dict):
        with open("logs.txt", "a") as f:
            f.write(f"{datetime.datetime.now()} | {endpoint} | {input_data}\n")

In your API route:

    log_usage("/generate", {"prompt": request.prompt})

---

10.7 Automation Tools (Advanced)

Tool	Use Case
cron + curl	Auto-ping API to avoid cold starts (Railway)
PostHog	Track frontend behavior/events
Sentry	Monitor frontend/backend errors
Google Analytics	Track public usage

---

10.8 Final Cost-Safety Checklist

Task	Done?
Prompt/input length limited	✅
Cooldown between API calls enforced	✅
.env keys secured	✅
Fallback error handling added	✅
Logs monitored and reviewed weekly	✅
Platform usage dashboards checked	✅

Bonus: Add a Usage Warning in UI

    if (usageCount >= 3) {
      alert("You’ve used 3/5 free generations. Upgrade for more!");
    }

Great for apps you plan to monetize later!

---

Chapter Summary

• You now know how to control AI API costs using simple tricks

• You can monitor app health and usage through logs and dashboards

• Your app is now safe for public sharing or demoing

---