Added Gemini TTS to My App and Lost Money in 24 Hours 🤡

(A small but painful lesson about AI billing)

Before this happened, my small TTS app was doing fine.

Not many users, but a few people were paying.

Nothing impressive, but profitable.

Then I thought:

“What if I add Gemini TTS to make it more premium?”

And that was when everything went wrong.

Background

My app is a simple web-based text-to-speech service:

1. reading documents
2. reading study materials
3. users paste text → get audio

Initially, I was using:

1. OpenAI TTS
2. Azure TTS
3. Google Cloud TTS

They all share the same characteristics:

1. pricing based on character count
2. easy to estimate costs
3. retries or errors are not financially dangerous

The app was simple, but it worked.

A few customers, no losses.

Then Gemini TTS Came In

To be fair, Gemini TTS is very good:

1. natural voice
2. smart pauses
3. good English and Vietnamese pronunciation

The demo sounded amazing.

So I pushed Gemini TTS into production.

What Happened?

After just one day:

1. Revenue: ~200,000 VND
2. Gemini TTS cost: over 400,000 VND

👉 Instant loss.

The most frustrating part:

1. user count didn’t increase
2. request volume wasn’t high
3. dashboard looked normal

But billing told a very different story 💀

The Problem Isn’t That Gemini Is Expensive

The Problem Is That You Can’t Estimate the Cost

After digging into it, I realized why.

1. Gemini TTS Does NOT Charge by Characters

It charges based on:

1. audio duration
2. pauses / silence
3. prosody (intonation, emphasis)
4. retries / reconnections

👉 Slow reading + pauses = much longer audio

👉 Longer audio = much higher cost

2. Dashboard ≠ Billing

The dashboard mostly shows:

1. request count
2. input tokens

But the real cost driver is:

1. audio output tokens

And that part is not clearly visible.

3. Retry Is Not Always Labeled as “Retry”

Cases like:

1. stream failure
2. connection drop
3. re-rendered audio

👉 Billing still counts them

👉 Dashboard doesn’t clearly say “this was retried”

Quick Comparison

OpenAI / Azure / Google Cloud TTS

1. priced by characters
2. 1,000 characters ≈ predictable cost
3. easy to estimate
4. very suitable for SaaS

Gemini TTS

1. priced by audio duration
2. audio length is unpredictable
3. silence and pauses cost money
4. very hard to control spending

👉 Amazing for demos

👉 Dangerous for production

The Irony

Before Gemini:

1. simple app
2. few features
3. still profitable

After Gemini:

1. no new users yet
2. already losing money

It really felt like:

“Adding a premium feature and accidentally shooting myself in the foot.” 😅

Lessons Learned

After this experience, here’s what I learned:

1. A powerful tool doesn’t always mean the right tool
2. For small SaaS products, predictable cost > wow factor
3. Billing model matters as much as model quality
4. New features don’t automatically create new revenue

Conclusion

I lost more than 400k VND, but luckily:

1. user base was still small
2. the issue was caught early
3. the app hadn’t scaled yet

If this happened with higher traffic,

it could easily have burned millions in a single day.

Now I’ve:

1. removed Gemini TTS from production
2. switched back to OpenAI / Azure / Google
3. limited character length per request
4. stabilized infrastructure costs

Gemini TTS is impressive,

but it truly is a double-edged sword.

And for my app, I chose to put that sword away.