The Hidden Cost of Using the Wrong AI Model

0:00 A major topic in AI this week is the soaring cost of AI spending and the problems that are emerging from this.

0:08 We see here in this article that Uber has burned its entire 2026 AI budget in just four months due to the massive amount of tokens used by Claude Code.

0:21 In this article here we see the CEO of Microsoft is telling his employees to use the appropriate AI models. He tells them not to use frontier models for non-frontier purposes.

0:32 And this shows how this issue of overuse is actually a problem of AI literacy, where people don't actually know the appropriate AI model to use in the appropriate case.

0:46 And so that's why in this video we're going to use this beautiful graphic created by Claude AI Design for us to break down the different models in ChatGPT today on June 19th, 2026.

1:00 before we break down all of the models and their token usage, I'd actually like to take a moment to explain what are tokens.

1:08 So, tokens are large language models' version of words. So in the English language, when I say, don't forget the human part, that is five words.

1:23 Models count these a little bit different, and so you'll see maybe punctuation or compound words Bye! Splits those into separate, into multiple words, so this is OpenAI's Tokenizer platform to help us understand how the model actually sees words in terms of tokens, so let's just test this phrase So,

1:46 already we can see DON'T is counted as two tokens due to the space that I placed after, so the DON'T counts as one token for the model and then the SPACE counts as a second token You can see down here the tokenization as I typed DON'T FORGET THE HUMAN PART Right now it's five tokens and five words What

2:09 happens when I add an exclamation mark? Yes, so we see that counts as its own token And you can see that for the longer words, it actually counts as multiple tokens, so let's test miss, M-I-S-S-I-S-S-I-P-P-I, Mississippi It's broken down, okay, only two tokens for that But you can see, roughly, that

2:34 tokens translate into words in our English language, but of course punctuation or multiple syllables or compound words can make it more than one word.

2:44 And different models have different tokenizers, So it might be slightly different depending on which eye you're looking at. But hopefully this gives you a little bit of intuition into, okay, how does the model actually interpret tokens?

3:01 So now I think we're ready to break down this diagram. So this shows us the four models that are within ChatGPT right now.

3:10 You have the Auto, Instant, Thinking, and Pro models. And automatically you're going to want ChatGPT is on this Auto account.

3:19 And this is where ChatGPT has its own internal router model, where it reads your input, and it decides which model is best to send it through.

3:30 And this is good for low-stake tasks that maybe you don't want to care too much about. However, if you want to be in control, it's best that you understand how these work, and you get to select and stay in control of which model is being used for your task.

3:45 So, the next model is this Instant model. This is great for 90% of tasks. It gives you quick answers. It's good for straightforward things like summarizing or a short explanation.

3:59 The next model that I think has a little bit more character to it is this Thinking model. And actually, right now, there's two versions of this.

4:08 You have the standard and the extended. Now, if you're doing a task that requires a lot of thinking, I recommend you be in the extended version of thinking.

4:20 Now, thinking works through a process called chain of thought. So, basically, instead of where instant will give you an answer directly, the thinking model is going to think about the process logically and develop a thought based on your answer.

4:37 So, say we have a question, a train leaves at 3pm, and arrives at 5pm.

4:55 How long did it take? Well, the instant model can immediately say 2 hours, while the thinking model, on the other hand, will calculate this step-by-step and very logically.

5:07 It'll say the departure was at 3pm. The arrival was at 5pm. Let me calculate the difference. This is 2 hours.

5:16 Let me verify that there was no date change. Let me be very precise about this. Now I can answer that the change, or the arrival time, was 2 hours.

5:25 And so, it really thinks through this, uhm, and that, of course, drives up the tokens. And so, if it's a task, such as calculate the time change, the change in time between 3pm and 5pm, you'll want to use this.

5:41 Because, ChatGPT has become quite capable. So, unless it's a task such as a business plan and you want it to research the competition, or anything, uhm, that involves real decisions with real effects, The thinking model is good because you really want it to think through each path and give you the

6:02 best developed answer. Now, the final model is this Pro, and this will be used, I think, very rarely. But any time you have to do, uh, scientific research or anything that requires a lot of technical expertise, Pro is your go-to model.

6:23 I would also recommend using this extended model, which uses a lot more thinking capabilities. Of course, the tokens. The tokens are a lot higher here, but yes, anything technical, the Pro model is your go-to.

6:36 Overall, I think most of your time should be in this instant and then in the thinking model. Unless, of course, you have a more technical career.

6:45 And this, of course, uses the deepest reasoning and the most amount of tokens. And so, each one of these I think of as a different digital colleague.

6:55 You want to be optimizing the task for the right colleague. And so, I love this learn-error rule. The first question you ask is, is it simple?

7:03 If this is true, you use the instant model. Does it need real analysis? If so, go with the thinking. And is the task high-stakes or research-grade?

7:14 If yes, use pro. And so, hopefully, with this informed input, you'll make the best decision, you can optimize your token usage, so you're not running your company through its AI budget in just four months,

7:29 and so you can have the most precision and efficiency for your tasks.