Monetizing AI-Based Apps With Usage-Based Billing

In the last year, we have witnessed a surge in AI-based applications and APIs following the introduction of ChatGPT. This sparked the emergence of new enterprise use cases and workflows. Many of these solutions are delivered as APIs rather than traditional UI-based applications and target developers.  

However, monetizing AI-based APIs demands a strategic approach to ensure that the cost of API access reflects the value provided to customers. It's crucial to mitigate the potential for misuse and ensure that the pricing model is in line with the infrastructure costs, which can be substantial for powering these AI-based APIs. As a result, the conventional SaaS pricing model, offering unlimited API access, is not suitable for AI-based APIs. In this article, let's delve into potential methods for monetizing AI-based APIs.

Ensure Pricing Aligns to Customer Value

AI-based APIs, such as those using large language models (LLMs), are typically contextual and generate a response from some input. However, different inputs can create drastically different results.

A single, complex request that generates a large amount of information is likely far more valuable to a customer than a request that provides very little information or none at all. Of course, a customer could make a large number of small requests to formulate a result similar to one large request.

It may, therefore, make sense to align your pricing with the results rather than API calls. For LLMs like ChatGPT, one way to do this is to base pricing on the number of input and output tokens. For an audio API, on the other hand, it may mean looking at the length of the audio clip in minutes, as it doesn't matter whether the customer made a single API call or one hundred API calls to generate the necessary result.

It's also important to consider the accuracy and validity of the result. If the result generated has low confidence—potentially due to hallucinations—then it's not as valuable to the customer as an accurate and complete result.

One way to do this is via pay-as-you-go billing, also called consumption-based billing, which can charge customers based on their consumption.

Ensure Cost Scales With Pricing

Unlike traditional SaaS applications, infrastructure costs for AI-based applications typically make up a larger percentage of the overall operational budget. This can include the computing cost to train a new model, the cost of consuming other APIs like OpenAI, and the hosting cost to serve your API, among other things.

As a customer consumes more of your APIs, your direct cost typically goes up as well. However, not all API calls cost the same. Different requests could have different unit economics depending on the request inputs, which model is used, etc. For example, if you are leveraging ChatGPT, a request with a larger number of tokens will cost more than a request with a smaller number of tokens.

Similarly, different models and contexts have different price points. Using the 128k context GPT-4 Turbo costs more than using the GPT-3.5 Turbo with 16k context. Because of this, you'd want to also ensure this cost difference is incorporated in your API pricing. Otherwise, you might have price arbitrage.

Prevent Accidental Abuse

Even though the cost to host an AI-based API is high, each API call can be seen as a commodity. A developer using your API likely does not require extensive set-up work just to start using the API. This can open you to abuse and arbitrage.

One way this can happen is with postpaid billing, which is pay-as-you-go. A customer could sign up, enter their credit card, and drive up a $10,000 bill. By the time the credit card is charged, however, it's canceled.

There are a couple of ways to prevent this:

Explore Prepaid Billing

Prepaid billing is one way to reduce the risk of abuse by requiring your customers to pay upfront.

This can be done by either requiring the customer to purchase a predefined quota or by purchasing credits that can be consumed over time. Quotas are perfect when the usage is predictable. However, customers of AI-based APIs may have a large variance in their monthly usage.

Prepaid credits make the most sense here. A customer could use a small number of credits in one month and a large number of credits the following month.

Consider Threshold-Based Invoicing

The tricky part with prepaid billing is it creates friction in the onboarding. You are asking a customer to purchase something before they are fully successful with the API and know how much they plan on using.

One way to solve this while still preventing abuse is with a combination of postpaid billing and threshold-based invoicing. With this setup, a customer is still invoiced in arrears.

However, the invoice is sent as soon as a threshold is met. For example, if the threshold is $500, then an invoice is sent and payment made as soon as $500 of consumption is made rather than waiting until the billing day. This model is popular within the advertising industry as the credit extended to a customer is capped to the threshold.