How Google's API Nearly Bankrupted My Side Project (And What I'm Doing About It)
-
The Wake-Up Call
Two days ago, I got a billing alert from Google Cloud that stopped me in my tracks.
Over 600 kr. In two days.
To put that in perspective: that's more than my entire revenue from PNL Pro for the past six months. Three days of premium TTS voices nearly wiped out the project.
That spike? 4140% higher than last month. Not a glitch (as much as I wanted to believe), just the reality of API economics I hadn't fully understood.
How did I get here? Let me rewind a few months.
The Journey
October 7, 2025 — I launched PNL Pro. I was excited. The first version only had cloud syncing, and honestly, only one friend seemed to care. But it was a start.
October 25, 2025 — The first real milestone: PNL Reader 2.0. I added a read-aloud feature using Google's TTS service. Users could now listen to web novels chapter after chapter, hands-free.
November 22, 2025 — The second milestone: TTS Pro for Dictionariez. AI-powered text-to-speech for language learners to hear sentences or paragraphs.
Both products used Google's Chirp3-HD voices — premium AI voices that sound remarkably human.
The feedback was great. Users loved the quality. I was proud of what I'd shipped.
For months, everything seemed fine. The bills were manageable. Then came March 2026.
I'd made a classic startup mistake: I optimized for user experience without fully modeling the unit economics at scale.
The Numbers
Here's what the billing breakdown revealed:
3,638,082 characters processed through Chirp3-HD in just a few days.
Google charges US$30 per million characters for Chirp3-HD. Standard voices? US$4 per million, 7.5x cheaper.
$30 per million sounds fine until you understand the use case. PNL Reader is designed for continuous reading. A user going through 50 chapters consumes ~250,000 characters per session.
The math:
- 250,000 characters × Chirp3-HD = $7.50 per session
- Same session with Standard voices = $1.00
- Power user reading daily for a month = $225/month for ONE user
One user's weekend of reading could cost more than their annual subscription.
With multiple power users consuming millions of characters daily, the costs scaled faster than I'd anticipated.
The Pivot
Once I understood the problem, I moved quickly.
Immediate action: Switched to Google Standard voices. Lower quality, but sustainable. This is temporary—at most until the end of this month while I implement better solutions.
What I'm exploring:
1. Local TTS. Browsers and operating systems have surprisingly good built-in voices now. Chrome, macOS, and Windows all support local TTS—free and faster than cloud APIs (no network latency).
2. Tiered pricing. Match the cost structure to usage:
- Free tier: Local TTS (free, fast, decent quality)
- Standard tier: Google Standard voices (affordable, good quality)
- Premium tier: Chirp3-HD or Gemini TTS (expensive, best quality)
Users who want premium voices pay for the premium cost. The economics align.
What do you think? Would you pay extra for premium voice quality? I'm genuinely curious. And your input will shape the product direction.
Looking for Alternatives
If you've worked with TTS services, I'd love your input.
What I know:
- Google Chirp3-HD (expensive but great)
- Google Gemini TTS (great but slow)
- Google Standard (affordable but basic)
What else is out there? ElevenLabs? Azure? Amazon Polly? OpenAI? Coqui? Tortoise TTS? If you've used any for long-form content, how's the quality, pricing, and latency?
Any recommendations would be valuable.
Key Takeaways
- Model your unit economics early. API costs can scale faster than you expect.
- Premium features need premium pricing. If the backend is expensive, the pricing should reflect that.
- Set up billing alerts. They won't prevent the problem, but they'll help you catch it before it's too late.
What's Next
The TTS feature continues with Standard voices for now. I'm working on better solutions and expect improvements by next month.
For developers reading this: model your API costs before you ship. That premium API might be amazing, but make sure the economics work at scale.
For PNL Pro users: thanks for sticking with me. Better audio is coming.
Want to support the project? Upgrade to PNL Pro.
Support me!