Massive Token Discrepancy – Vapi Key vs. Own OpenAI Key – Unjustifiable LLM Overbilling
Hi Vapi Team,
We’re actively using your GPT-powered phone assistant service with a fully token-optimized setup – streamlined prompts, highly compressed tool context, and minimal turn count. However, we recently uncovered a massive and reproducible discrepancy in token usage between your OpenAI key and our own key, even for identical calls.
We ran the same test call twice:
Cost Breakdown – Discrepancy Example
| Description | Prompt Tokens | Completion Tokens | LLM Cost (USD) | Notes |
| --- | --- | --- | --- | --- |
| Vapi Key | 21,586 | 42 | 0.0435 |
Wildly inflated |
| Own OpenAI Key | 10,811 | 40 | 0.0000 |
As expected |
We’re actively using your GPT-powered phone assistant service with a fully token-optimized setup – streamlined prompts, highly compressed tool context, and minimal turn count. However, we recently uncovered a massive and reproducible discrepancy in token usage between your OpenAI key and our own key, even for identical calls.
What we did
We ran the same test call twice:
- Once with **your Vapi OpenAI key**
- Once with **our own OpenAI API key**, so we could monitor usage on OpenAI’s dashboard directly.
Both calls used:
Identical prompts and context
Same toolkits
Identical user interaction
GPT-4.1 (gpt-4-turbo)
Call IDs
- **Vapi Key Call:** x-call-id: a847c3e6-cfbc-497f-bf62-21e397a32d95
- **Own Key Call:** x-call-id: 4f004ae9-9149-4736-a13f-a29b5eaed245
Cost Breakdown – Discrepancy Example| Description | Prompt Tokens | Completion Tokens | LLM Cost (USD) | Notes |
| --- | --- | --- | --- | --- |
| Vapi Key | 21,586 | 42 | 0.0435 |
Wildly inflated || Own OpenAI Key | 10,811 | 40 | 0.0000 |
As expected |This is not just 2x – the discrepancy escalates with longer calls, growing exponentially the more tokens are exchanged. This is just the smallest call example.
