[INFO] OpenAI Costs

[pmterp]

Would anyone be willing to share estimates of their OpenAI costs related to transcription and analysis? We're beginning to use it internally, but we serve a wide variety of clients (sizes and industries) and would love to be able to ballpark the estimates. I'll be happy to come back and share our findings as we get them.

 

[nikolascx]

Hi Happy new year
For which service?
Is it for Transcription? For AI Agents?
The calculations and prices are different.
This also depends on which models you are planning to use.
You can start from here https://openai.com/api/pricing/

 

[KyriacosS_3CX]

I would also suggest that you do a few practical tests, have a few calls go to OpenAI, note their duration etc, and see their breakdown of resources used and cost - if you know how many tokens a certain action/call with a certain duration costs (*words might be a better metric than time), you can scale that up and down accordingly, and also be able to adjust cost estimates if the per token price changes with time/model use.

 

[pmterp]

Thanks guys. For this post, I'm specifically referring to transcription. We've now setup our OpenAI account and will start testing and getting some ideas. However, I was hoping to get some real-world examples of what others are seeing.
For example, a small install with 15 users, all with transcription enabled logged x number of calls and used x number of minutes and the associated OpenAI cost was around $X

Again, not looking for specific numbers that I can quote a customer but at least some ranges. I can say "Other businesses similar to yours have spent between $400 - $500 per quarter in OpenAI costs."

However, now that I ENT+ is an option, we'll probably start by recommending that route. If the client has other AI needs like agents, this info could be helpful.

 

[nikolascx]

Your question is like asking me "What is the bandwidth required for a small company of 15 users?"
What if you have 1 "special user" that every morning decides to perform a special routine? - then all your calculations will be skewed.

To play along with your scenario, (trying to give you something useful in the meantime), lets calculate on 1 user, making 1 call, of 1 minute and 1 call of 1 hour. Assuming a bunch of stuff in the meantime.

Part 1: The transcription
The cost of openAI's model to transcribe is $0.006 per minute. (Standard rate)
If 1 user, makes 1 call, of 1 minute duration, the cost will be $0.0060
If 1 user, makes 1 call of 60 minutes duration, the cost will be $0.3600

Part 2: The Reasoning
Then you have the reasoning or summary this depends

1. on the model selected (there are around 6),
2. the speech rate
3. Whether the utterances are simple or complex. (Simple means less tokens - complex means more tokens).
4. It also depends on whether cached tokens are used or not.

Assumptions

• For this scenario, Ill select the most expensive model gpt-4o model,
• 2 perfectly normal people to have a perfectly normal conversation at a normal spoken speech rate.
• The number of input tokens should be around 600 tokens
• I will also assume that the number of output tokens = number of input tokens so 600 IN 600 OUT
• Which means that 1 minute of audio = 600 tokens and 1 hour of audio = 36k tokens

Rate
Input tokens: $2.50 per 1 M tokens
Output tokens: $10.00 per 1 M tokens
600 tokens = 0.0006 M tokens

Calculation
Input cost: 0.0006 × $2.50 = $0.0015
Output cost: 0.0006 × $10.00 = $0.0060
Total GPT-4o cost 1 minute (input + output): $0.0015 + $0.0060 = $0.0075
Total GPT-4o cost 1 hour (input + output): $0.0075 * 60 = $0.45

=========
Total price Price of Part 1 (Transcription) + Price of Part 2 (Reasoning)
For 1 minute duration, the cost will be $0.0135
For 60 minutes duration, the cost will be $0.81

I hope you find this informative and make sure to make calculations with other models and correct token rates.

 

[pmterp]

=========
Total price Price of Part 1 (Transcription) + Price of Part 2 (Reasoning)
For 1 minute duration, the cost will be $0.0135
For 60 minutes duration, the cost will be $0.81

I hope you find this informative and make sure to make calculations with other models and correct token rates.

Thank you, that is very detailed and helpful. I completely understand that we can't "quote" off these numbers and the variables are extensive.
Once we get a few clients using it for a month, I'll be happy to report back with estimates including the number of users, calls, minutes, and costs. I feel like it would help establish some examples that could be used when approaching potential customers.

 

[N_G]

If you want to use a package and know exact cost then just use ENT+ which gives you an exact price per year within fair usage and you also get margin on it.