Chief Evangelist and VP of Innovation at Pricefx. Brand ambassador, keynote speaker, author, pricing expert & podcaster.
OpenAI’s ChatGPT set records for an application’s adoption, hitting 100 million monthly active users in just two months. It has sparked the imagination of the masses and kicked off an AI gold rush.
Major players such as Amazon, Microsoft, Google and Salesforce, as well as startups like Anthropic, Cohere and Dataiku, are building solutions to enable businesses to take advantage of AI across various use cases. While there is no doubt that generative AI is a game-changing technology, I find that the pricing model for these products is less than stellar.
The Need For Transparent Pricing Models
First, a great pricing model is clear and value-driven. This means that pricing is transparent, easy to understand and based on what drives value for the customer.
Cheekily, I asked ChatGPT if its own pricing model was clear and value-driven. It seems to be aware of some of the shortcomings, telling me that although “OpenAI’s pricing is competitive within the market and offers a high-quality service… some users may find the pricing model confusing or too expensive for their needs.”
On LinkedIn, I asked other pricing professionals what they think of OpenAI’s pricing model. The results were mixed, but the majority of people I asked described the model as unclear and not value-based.
Looking At Current Pricing Models
Using one of the largest organizations in generative AI, OpenAI offers two pricing models for its GPT-4 language model API: usage-based pricing and subscription-based pricing. Usage-based pricing charges customers based on the number of API requests made, while subscription-based pricing offers unlimited usage for a fixed monthly fee.
At the time this article was written, OpenAI’s usage-based pricing starts at $0.0075 per token for the first 100,000 tokens, with decreasing prices for higher volumes. The subscription-based pricing starts at $399 per month for the developer plan, which includes up to 5 million tokens per month, and goes up to $12,000 per month for the team plan, which includes up to 500 million tokens per month. There are variations depending on the different models for text, chat and embedding.
In essence, this is a usage-based model that is based on the underlying costs to run the AI on GPUs. One of the reasons I believe people feel it is not clear is because it is difficult to translate into a price for a given piece of work. Furthermore, as it primarily relies on the cost-to-usage ratio, it does not directly correlate with a value-driven outcome. Compounding this issue, there are many different prices per model.
Compare this to Google’s pricing model, which is based on characters and only has one model/price per category of text, chat and embedding; it is much easier to understand, but it is still entirely based on cost per usage. It also appears to be more expensive than OpenAI for most use cases.
The Cost Behind The Technology
In order for a large language model (LLM) to run and produce results, it requires graphics processing units (GPUs). While originally developed for computer and video game graphics, their ability to offer parallel processing has made GPUs the engine for AI applications.
GPUs are expensive to buy and run, as they consume a lot of energy. The token-based model employed by OpenAI is actually tied to this idea of usage and the inherent cost of the underlying infrastructure that powers LLMs. This is true of Google Cloud Natural Language API, Amazon Comprehend and IBM Watson Language Translator. In short, a cost-based pricing model is the most common for AI language models.
For the buyer, this can be challenging. The current pricing model is also usage/consumption-based. The more you use it, the more you pay. However, LLMs can be used in a large range of applications and it’s not always clear how end-users will engage with the application, making pricing subject to sticker shock.
At the same time, the value drivers vary a lot by application, so a value-based approach is difficult to apply in practice. But buyers do not care about the cost per second per GPU; they care about outcomes and the value they drive. There are some tools that OpenAI has put out to be able to enter text into and understand the cost for smaller jobs, but this can be a challenge for larger jobs and from a budgeting perspective.
Improving Pricing Models
For generative AI’s pricing models to improve, I think they need to be based on the segment of the customers, their application of this technology and the value it drives for them. This understanding of customers and the value drivers—the ability to segment them and differentiate prices accordingly—is the foundation of great pricing models.
Generally, the alternative for using these services includes people or other generative AI models, so these value and market factors should be incorporated into the model. A simplification of the pricing tiers and models would help. It should be easy to determine for a given piece of work how much it will cost, and ideally, this should be a maximum so that customers can budget appropriately.
It will be interesting to see how this develops as these models mature and expand. I hope this article can help people understand the current landscape and also provoke some thought as to how to simplify and make the pricing models more value-based in the future.
Forbes Business Council is the foremost growth and networking organization for business owners and leaders. Do I qualify?
Read the full article here