OpenAI's business model starts to take shape
With the release of an enterprise-grade API and the fine-tuning version of GPT 3.5-Turbo, OpenAI's business is starting to come into shape. And so will other API providers.
OpenAI in the past week or so has started to make its business model a little clearer as it starts to take shape as a tiered pricing structure with increasing premium performance and accessibility.
With the release of fine-tuning for GPT 3.5-Turbo, and the announcement of a new Enterprise tier, its business model is starting to take shape. After launching the GPT 3.5-Turbo API for developers, OpenAI is starting to offer increasingly premium options to fill performance tiers for multiple use cases.
From a product perspective, OpenAI’s dual announcements more or less come down to rate-limiting. Many companies I’ve spoken with have wanted to use GPT-4 due to its drastically better performance than GPT 3.5-Turbo, but run into issues with rate limiting and reliability.
As all these pricing schemas come out, we’re starting to see OpenAI the business start to take shape around ten months after OpenAI the product came out. GPT-3 has been out for a considerable amount of time, but OpenAI’s powerful Chat Completions API tools seem to be what will be the primary driver of its growth as it starts to build out an enterprise strategy.
And as OpenAI leads, so others will likely follow. The foundation model providers are all racing to acquire the best hardware en masse, raising hundreds of millions of dollars to keep up with one another. They’ll amass largely similar massive training sets, and while there might be differences in curation and technique, they’re all hoping to be the kind of backbone of an API-delivered generative AI mesh—at least, until one of them decides to launch a product that can plug into a VPC.
OpenAI’s announcement for its enterprise structure positions itself around offering benefits for privacy and compliance, but at the end of the day it’s about performance. Specifically this kind of performance:
OpenAI is basically releasing a preferred GPT-4 tier to go alongside its normal GPT-4 tier, which is really powerful but has its issues. A tiered structure like that generally carries an up-and-to-the-right pricing system, which you can see in the availability of its APIs—and each one fits neatly into four or five boxes.
Reviewing OpenAI’s cost structure
We’re seeing all of OpenAI’s “product lines” basically synthesize into four (or five, depending on your perspective) different usage tiers.
Base: The standard GPT 3.5-Turbo and Davinci models, which is for at-scale operations that don’t require a ton of accuracy and quality. (This is probably closer to text extraction or basic completion.)
Advanced: GPT 3.5-Turbo’s fine tune option, designed for at-scale operations that require more accuracy or better performance for specific use cases. (This would probably be around chat bots.)
Limited-use GPT-4: This tier seems to give companies access to a “weaker” GPT-4 API which carries some rate-limiting issues. (This one’s probably useful for one-off jobs that need the quality of GPT-4 but aren’t in active deployment.)
Production-grade GPT-4: This tier is in place for companies that need the performance and quality of GPT-4 at scale without restrictions and with guarantees around reliability. (This probably goes into actual live products with considerable usage.)
You could add a “free” tier here for the base use of ChatGPT via a web interface, which would essentially come down to employees copy/pasting information into it. And honestly it’s pretty likely employees (particularly at startups) are already doing this absent of the blessing of an organization’s IT department in the same way they were with Dropbox/Box/etc. back in the early 2000s.
Here’s a rough cost breakdown of the existing APIs (with PaLM2 on Vertex AI thrown in there for funsies) per 1,000 tokens. I’m not including GPT-4’s 32K context window on this one, which is a complicated subject that we’ll address at some other time: