Clone OpenAI pricing
Replicate OpenAI’s per-token pricing model with Lago.
In this article, you will learn how to build a billing system with Lago based on tokens. This template is suitable for Large Language Model (LLM) and Generative AI companies whose pricing can vary based on the application or model used.
Summary
Aggregate usage with filters
Use single metric to meter tokens with different filters. (More here).
Set up a per token pricing
Create a plan to price packages of tokens used. (More here).
Ingest usage in real-time
Retrieve consumed tokens in real-time. (More here).
Pricing structure
For OpenAI, pricing depends on the language model used. Here are several price points they offer: “Prices are per 1,000 tokens. You can think of tokens as pieces of words, where 1,000 tokens is about 750 words (learn more here).”
GPT-3.5 Turbo pricing
| Model | Input | Output |
|---|---|---|
| 4K context | $0.0015 / 1,000 tokens | $0.002 / 1,000 tokens |
| 16K context | $0.003 / 1,000 tokens | $0.005 / 1,000 tokens |
GPT-4 pricing
| Model | Input | Output |
|---|---|---|
| 8K context | $0.03 / 1,000 tokens | $0.06 / 1,000 tokens |
| 32K context | $0.06 / 1,000 tokens | $0.12 / 1,000 tokens |
Step 1: Aggregate usage with filters
Lago monitors usage by converting events into billable metrics. To illustrate how this works, we are going to take GPT-4 as an example. OpenAI’s GPT-4 pricing includes a single metric based on the total number of tokens processed on the platform.
To create the corresponding metric, we use the sum aggregation type, which will allow us to record usage and calculate the total number of tokens used. In this case, the aggregation type is metered. This means that usage is reset to 0 at the beginning of the next billing cycle.
For this metric, there are two dimensions that will impact the price of the token:
- Model: 8K context or 32K context; and
- Type: Input data or Output data.
Therefore, we propose integrating these two dimensions into our metric as filters:
- Filter #1: Distinguishes between various models utilized; and
- Filter #2: Separates input and output types. By implementing these filters, we can assign distinct prices to a single metric, based on events’ properties.
Step 2: Set up a per token pricing
When creating a new plan, the first step is to define the plan model, including billing frequency and subscription fee. OpenAI pricing is ‘pay-as-you-go’, which means that there’s no subscription fee (i.e. customers only pay for what they use).
Here is how to set the monthly plan for GPT-4. Our plan includes the ‘per 1,000 tokens’ charge, for which we choose the package pricing model.
As we have defined 2 filters (models and type), we can set a specific price for each Model/Type combination.
We can apply the same method to create plans for GPT-3.5 Turbo. Our plan is ready to be used, now let’s see how Lago handles billing by ingesting usage.
Step 3: Ingest usage in real-time
OpenAI records the token usage, the number of images, and the usage of transcribing speech into text. These activities are converted into events that are pushed to Lago. Let’s take GPT-4 as an example: Lago will group events according to:
- The billable metric code;
- The model; and
- The type.
For each charge, the billing system will then automatically calculate the total token usage and corresponding price. This breakdown will be displayed in the ‘Usage’ tab of the user interface and on the invoice sent to the customer.
Wrap-up
Per-token pricing offers flexibility and visibility, and allows LLM and Generative AI companies like OpenAI to attract more customers. With Lago, you can create your own metric dimensions to adapt this template to your products and services.
Give it a try, click here to get started!
Was this page helpful?