Calls not charging the stated standard pricing
# supporth
Huza
04/14/2024, 4:36 PMHey i am noticing that the cost per minute is not predictable whatsoever on my side
For the same assistant using the exactly same settings, the cost per minutes varies way too much. Additionally, its almost never as low as what its calculated on the dashboard screen.
Calculated: 0.16/ min for GPT 4, 0.10/min for 3.5
Actual costs when testing:
ID
78f44e51-7752-4ab0-903c-306d0e07c425
0.33 / min
ID
bdc8bf0c-1336-4459-a4d1-683e334cdfa3
0.35 / min
ID
9ea9e0e0-a521-49b9-87b9-40fbf1f84df3
0.18 / min
ID
459e0165-e200-4c2d-9625-e749c82f7b54
0.35 / min
ID
fcfdaebd-838b-45bd-b58c-a9418cdf565e
0.43 / min
The cost is mostly on average double or triple the shown pricing, and at times spikes to quadruple.
I haven't connected my own Accounts to Vapi so i am not sure if this might be a factor, but i will appreciate understanding what causes this and if it is controllable
v
Vapi Support
04/14/2024, 4:36 PMHi @Huza, I'm an AI assistant for Vapi. While a team member reviews this question, I'll look through relevant sources and see if I can provide suggestions. Please tag @Vapi Support with any follow-up questions.
Vapi Support
04/14/2024, 4:37 PM### Answer
The variability in the cost per minute you're experiencing could be influenced by several factors not directly visible in your dashboard's estimated calculations. According to the Vapi pricing documentation, the platform bills for Transcription and Voice services at cost, and if you use a Large Language Model (LLM) through Vapi, you're billed for the Model at cost as well [(1)]() . Additionally, Vapi charges a base fee of $0.05 per minute on top of provider costs.
If you haven't connected your own accounts to Vapi using Provider Keys, you're charged directly through Vapi at these rates, which might not reflect any special pricing or discounts you could have directly with the providers [(2)]() . This could be a factor in the discrepancies you're noting.
Moreover, the cost can vary significantly based on the combination of services used per call. For example, different voice providers and models have different rates per minute, which can contribute to the overall cost variance you're observing [(1)]() .
To help manage and predict costs more accurately, consider:
1. Reviewing the detailed breakdown of each call to understand which components (Transcription, Model, Voice) are contributing most to the cost.
2. Connecting your own provider accounts through Provider Keys to potentially benefit from any special pricing or discounts directly from the providers [(2)]() .
Understanding these elements might help you better control and anticipate the costs associated with using Vapi.
Vapi Support
04/14/2024, 4:38 PMGot it. Tagging @User for additional help.
s
Sahil
04/14/2024, 5:18 PMHello Huza,
The cost largely depends on the amount of input and output text sent from the LLM model. In my experience, I was charged $1.38 per minute because I was retrieving the top 50 similar searches from Pinecone. This significantly increased the input and output text, thereby raising the cost.
I suggest experimenting with the input and output tokens to see if any modifications can reduce your cost per minute.
h
Huza
04/14/2024, 6:38 PMThat makes sense
I thought there would be a standard rate for LLM response fetching
I am not really using Pinecone so I'll ask, is there any adviced way to make this more predictable?
Huza
04/14/2024, 6:39 PMAnd the weird thing is
In some calls, the LLM cost is as low as 0.01/min
Where as in others, it goes as high as 0.27/min
Like in both, it IS having the same conversation
Huza
04/14/2024, 6:39 PMSo it just strikes me as odd as to what could be that massive of a variation
n
nikhil
04/14/2024, 6:44 PMIf you see call.messages, that should give exactly what went to the model (hence the token cost)