Which ChatGPT to Choose - 3.5, 4, 4.5 (0613, 0125, 1106)???

A year ago, ChatGPT pretty much came in one model, the ChatGPT 3.5, however newer models have since come out (with variations within the models), AND while you'd probably want the latest and greatest GPT 4 Turbo / GPT 4.5, it's helpful to know why as the older GPT 3.5 may be more applicable for specific use cases you have.

OpenAI vs Microsoft Azure OpenAI Variations vs Microsoft Copilot, etc

When you buy a "product that uses OpenAI" like Microsoft Copilot or other ready built solutions, you get what you get, and don't have the option of choosing which version of ChatGPT (3.5, 4.0, etc) it runs. It's the reason you pay a fee per person per month is to get something that does what it does and you don't have to think about the inner workings.

However, as I get further into this article, you might want the "latest and greatest" version of ChatGPT model to work your data, and that's where the flexibility of setting up your own Microsoft private AzureAI/ChatGPT comes into play.

Even OpenAI simplifies it a bit where you can buy the basic service (or use the free version) that uses the older/basic ChatGPT 3.5 model, OR you can pay OpenAI more $$ for their more current service running ChatGPT 4 or 4.5. So you pick your subscription, and you get the model you pay for.

THIS is why the "build your own" Microsoft AzureAI ChatGPT solution has been extremely popular for enterprises, because for 1 monthly price (~$350/mo) for your entire organization, you can build a private instance of Microsoft's OpenAI/ChatGPT setup, AND within that single instance you can slice it up and run multiple various models (3.5; 4; 4 Turbo; 0613, 0125, 1106, etc) side by side. I'll cover more on the variations and which use case fits the various models in the balance of this article.

Why ChatGPT 3.5, 4.0, 4.5, etc?

Without getting into a lot of technical details and focusing more on the usage results, bottomline is ChatGPT 3.5 is great for a search engine replacement in finding text, pointing to citations (reference documents), answering questions as a ChatBot. It's fast, cheaper to ask it thousands of questions than GPT 4, and for mundane ChatBot questions, it's plenty.

ChatGPT 4 has a couple notable improvements. ChatGPT 4 looks at more content at the same time before answering a question AND it solves complex problems better because of it. In a general comparison, the basic ChatGPT 3.5 model looks at a paragraph or two at a time, whereas ChatGPT 4 looks at 2-3 pages at a time. Because the "window of view" is bigger in GPT 4, it can compare and contrast content better, it can formulate answers in a page format than basically at a paragraph format. (I'm simplifying this a lot, but "seeing" and "processing" bigger blocks of data in GPT 4 makes its answers broader and provides a bigger picture view of things).

OpenAI has a GPT 4.5 model that does math, logic, and reasoning better than previous versions, which Microsoft picked up this release this past month which Microsoft calls it GPT 4 Turbo. For orgs that are working with financial spreadsheets, data tables, numeric data, this Microsoft AzureAI GPT4-Turbo is the better model to choose.

Newer / Bigger Model - Higher the Cost

For this, I'm going to swing solely to the Microsoft AzureAI solution as other services don't provide the flexibility to run multiple models side by side without requiring you to buy/replace licenses with completely different monthly costs. As noted earlier in this article, from your ~$350/mo to Microsoft, you can setup and run GPT 3.5, GPT 4, GPT 4 Turbo, etc all at the same time. If you get bored, you can load up multiple models with the same data and ask questions and compare the results.

We actually started this process to do a side by side comparison, we built out this beautiful table comparing the results between GPT 4 Turbo, GPT 4, and GPT 3.5, and after 4-5 sample sessions using the same questions against the various models, there was such variability of the nuances that we determined it wasn't worth publishing the comparative results beyond what I'll summarize later in this article. It was VERY subjective, at one moment we'd observe one model answering a question better than another model, but then you ask the same question again and it did a better job. But then the better model did worse, except when you asked it again and it did a great job. (its artificial intelligence, it'll provide slightly different responses every single time, VERY hard to qualify the specific differences)

From a cost basis, Microsoft has a pretty fixed cost at about $280/month to run AzureAI for an entire organization (with hundreds or thousands of users asking the system questions). The variability is in how many questions are asked in a month. If you've ever looked at Microsoft's pricing, this is that $0.002 per 1,000 tokens, blah blah blah. In basic terms, you can ask 4-8 questions for 1-penny, and then if you spend a dollar, you'll get in 500-700 questions. I always tell people it costs $350/mo to run this thing, which will cover 50,000 questions a month. That comes to 5 questions a second during business hours, which is a LOT of questions (or maybe not), this is why its important to setup and run a pilot to see how "your mileage may vary", usually falling on the lower cost side than higher cost side of things.

With A.I. usage, initially there's a lot of excitement to ask a whole lot of questions (10, 20, 30 questions) but then once the novelty wears off, then the business questions stabilize down to focused things that trickle in a few here/there. It's like the company Intranet, when it first goes live, everyone pokes around, looks at what's there, but after a while they only ask it a few questions a year on specific forms, documents, policies they might be interested in. A.I. tends to have more usage than an Intranet, but the same volume shift applies over time.

But this is where at some point you might consider using the cheaper ChatGPT 3.5 for basic questions/chats as ChatGPT 3.5 runs half the cost for questions and answers. Separately you can setup a GPT 4 Turbo instance for your analytical use case. Again, the various models can run side by side in a single $350/mo AzureAI instance, suited to your needs.

Variations 0314, 0613, 0125, 1106, etc

And to confuse it more, in the Microsoft AzureAi setup, your GPT 4 can be configured using an older 0314 variant, a mainstream 0613 variant, or one of the latest 1106 or 0125 variants (where 0125 is the latest and greatest at the time of this writing (Mar/2024)).

We've found the 0125 runs faster than the 1106 variant, and as of March/2024, Microsoft has retired all 1106 deployments and auto-shifted everyone to 0125, which is great. No modifications or changes required, and it runs faster.

So Which to Choose????

While you could go a bit nutty comparing the variations (GPT 3.5 vs 4, or 0613 vs 0125), I'll net it out to make it simple:

• Simple Chats / Simple ChatBot / High Volume / Desired Lowest Cost - Go with AzureAI/ChatGPT 3.5. It's half the cost for questions, and if the questions/answers are like old fashion "search" questions but insetad using this modern A.I. technology, this will get the job done
• Complex Chat / More Analytical Responses Desired - If you have more complex questions, want the best answers, go with the latest AzureAI/ChatGPT 4 0125-Preview that's available.

Monitor the costs to see if you're pushing your $350/mo budget estimate, or if you're good with $380, $400, or more in a month OR you just want the best ChatGPT model even if it costs a little more, then just build with the latest and greatest model for everything. Whole purpose of this article is to give you insight of the options.

It's literally fractions of pennies here and there, and again after the initial novelty wears off, the actual business uses of ChatGPT stabilize to specific measured tasks. Eventually you'll want to tie your AzureAI into doing data modeling, forecasting, analytics, maybe even image recognition, and that's when you are building on top of the basic ChatGPT to expand your business usage of the technology.