@Phone systems NZ
You have been kind to many on this forum - Thank you
Let's dive into this quick
Recommended starting point: Set your TPM limit at
20,000, then gradually increase to
40,000 once you’re comfortable.
(The exact ceiling you’re allowed to set depends on the tier OpenAI has placed you in. Higher tier = higher possible limit. I’m setting that aside for now, but it’s worth knowing it exists.)
Beyond the tier, token usage depends on:
- Language
- Model
- Conversation style
- Number of words/utterances
Here’s a realistic breakdown based on real-world usage:
Base assumptions
- Natural speech = 150 words per minute
- Maximum call duration = 7 minutes
One speaker: 150 × 7 =
1,050 words
Two speakers (balanced conversation):
2,100 words total
Tokens conversion (English) OpenAI’s widely accepted rule of thumb:
1 word ≈ 1.3 tokens 2,100 × 1.3 =
2,730 tokens
So a normal, balanced AI voice call typically lands between
2,500 – 3,000 tokens.
Now there is another factor -
the nature of the call.
- If Im calling to simply checking in or to schedule a meeting, thats easy stuff and the AI is almost only listening. So I would put this at 600 - 1000 tokens
- If I am having a balanced conversation about work, or reviewing a document, general chat etc, that would be around 2500-3000 tokens.
- If the call is dense and technical, I would estimate at 4000 tokens because of the heavy terminology.
- And if you happen to have knowledge base, add another 1500-2000 tokens overhead for the retrieval cost.
Conclusion:
A typical 7-minute production call with knowledge base ends up around
5,000 tokens.
That works out to approx
714 tokens per minute. For safety and buffer, I round this up to
1,000 tokens per minute per active call.
At 20,000 TPM ÷ 1,000 TPM per call =
20 simultaneous calls to the AI Agent..
Hope this helps
God bless
