Voiceflow : How to pick models and minimize costs 💡

Hey everyone!

If you are a beginner and want to be efficient as possible with your credit when it comes to using voiceflow agents according to their credit system, Here are voiceflow's guides and suggestions:

> What uses credits?

Messages
LLM usage
Voice call minutes
Text-to-speech

> What doesn’t?

Functions (no credit drain!)
Knowledge base queries

> Monitor usage with Analytics:

Use the credit usage chart to track what's eating your credits.
Check the prompt usage chart to see which prompts get triggered most.

🧠 Choosing the Right AI Model:

It's all about balancing cost, capability and latency when it comes to picking the right model for the task. It can be tricky — especially for voice-based agents. Here’s a general breakdown of LLM types:

Lightweight & Cheap (e.g., GPT-4o mini, Claude Haiku) : Great for basic tasks, fast and low-cost.
Balanced & Smart - More capable (e.g., GPT-4o, Claude Sonnet) : Ideal for reasoning, KB-heavy workflows, and accurate step-by-step instructions.
Most Powerful - Heavy (e.g., GPT-4, Claude Opus) : Best for tough questions or deep reasoning — but expensive!

❗Voiceflow recommends not using legacy models such as GPT-3.5 Turbo for building agents.

🧪 Test & Experiment

When switching models:

Test thoroughly
Monitor transcripts
Iterate & scale what works

Go to the project -> analytics when it comes to monitoring your agent.

Cheaper ≠ better. Choose intentionally. Some steps are worth the higher cost for better quality.

🎁 Bonus: Free Proactive Messages

Ever wonder how to make your webchat widget pop with a proactive message (like “Hi there! Need help?”)? Good news — proactive messages are free and don’t cost credits, since they exist outside your project logic.

I attached the code for the proactive messages provided by voiceflow.Make sure to wrap it with

setTimeout(function() { your chat widget code}, 350). Otherwise it won’t display due to widget loading delay.

I've been struggling with these questions myself. For anybody in the position I was hope this helps!!

0 comments