¶1The best people in the world at using AI don't care as much about prompt engineering anymore. What they do care about is which tokens they're using for which task. Not all tokens are made equal. And I'm going to explain exactly how to increase the quality of your AI usage while also decreasing the cost. This is what the absolute frontier of AI users are doing and I'm going to show you how they do it.
¶2And by the way, if you like when I explain how to get the most out of AI, be sure to like this video, subscribe to the channel, it really does help. All right, so first, what is a token? We're going to go over this quickly. A token is the basic unit of a large language model like ChatGPT or Claude. You can kind of think of them like words.
¶3So, let me just show you. So, this is the tokenizer that ChatGPT uses. That means how to take text and make them into tokens. So, here's an example. The cow jumped over the moon.
¶4Now, down here, each one of these colors is the token. So, you can see there are six total tokens, one for each word. But, full words aren't always just a single token. Sometimes, they're more. Here's an example.
¶5The word programmatic is actually two tokens, one for program and one for matic. So, that's how to think about tokens. And the way that large language models work is by predicting the next token. So, for example, the cow jumped over the the model hopefully would predict moon. And now, here's the big secret.
¶6Not all tokens are equal. Some are higher quality, some are faster, some are more expensive. And you want to use the right mixture of those factors for the specific task that you have to complete. So, when a new model gets released, I think a lot of people get confused because the first thing they do is go to the pricing page and look how much are these tokens. And of course, OpenAI and Anthropic have some of the most expensive models on the planet.
¶7Here's an example, GPT 5.6 Soul, which is the most recent release by OpenAI and also their most powerful model, comes in at $5 per million input tokens and $30 per million output tokens, which is quite expensive. Fable from Anthropic is even more expensive. And just last week, a new Chinese open-source AI model dropped called Kimi, and that came in at $3 per million input tokens and $15 per million output tokens, about half the price of GPT 5.6 Soul. So, you might be thinking to yourself, if they're scoring the same on different benchmarks, then you're essentially paying half the price for the same level of intelligence. But it's not that simple.
¶8Now, let's look at some of those benchmarks in isolation. We see Deep Seek, Frontier Seek, Kimi Code Bench, Terminal Bench. These are all very standard benchmarks that the Frontier Labs use. And as we can see, Kimi K3 is extremely competitive with the best of the best models coming out of Anthropic and OpenAI. But it comes in at half and sometimes less than half the price.
¶9Now, remember I said how not every token is created equally? Well, here's why it matters. It doesn't really matter how much the tokens cost. What matters is what you can actually get done with those tokens and a fixed price. So, let me show you a chart that explains what I'm talking about.
¶10This is artificial analysis, and this is the actual cost to complete tasks on average on their benchmark. So, what we're seeing here is Kimi K3 costs 95 cents per task completed, and GPT 5.6 Soul costs a dollar and four. Basically, about the same, very close. Now, when we come up to Claude Fable 5, we see it's $2.75. Very expensive.
¶11Now, keep in mind, this is for the same exact task. Then, if we scroll down a little bit more, we see something very similar. The cost to complete the entire benchmark. We're seeing 2,700 for Kimmy, we're seeing 2,800 for GPT 5.6, and 5,600 for Fable. So, let's think about this math for a second.
¶12We have GPT 5.6 coming in at $5 per million input, $30 per million output. We have Kimmy K3 coming in at $3 per million input, $15 per million output. And that means Kimmy is effectively half the price. But, is it? Because from what we've seen, it takes twice as many tokens for Kimmy to think through and solve a specific task than it does GPT 5.6.
¶13And that is what I mean by not all tokens are created equally. So, even though it's half price, what we're actually seeing is 2x tokens. And because it's half price, plus takes twice as many tokens, that is effectively the same price overall to accomplish a task. So, let's say 5.6 takes 1 million tokens to solve a task, that's $30. And Kimmy K3 to solve that same exact task takes 2 million tokens, that's also $30.
¶14And this intelligence density is extremely important. How much intelligence are you squeezing into that individual token? And right now, at least from what I've seen and in my opinion, OpenAI and ChatGPT are number one for token density. It takes Fable many more tokens to solve the same problem as GPT 5.6. Now, the next thing I want to talk about is very similar to not all tokens being equal.
¶15It turns out different models approach problems differently and they behave differently. They write code differently. They excel at certain things and they're worse at other things. And a really great technique to get the highest quality code output is letting two models essentially review each other. And it turns out the sponsor of this video Greptile just released a feature that does exactly this.
¶16Each model finds more bugs in the other models code. So in orange we're seeing Opus 4.7 in purple we're seeing GPT 5.5 and when Claude in orange authored the PR and then GPT 5.5 reviewed it, the GPT 5.5 actually found more bugs than if Claude just reviewed its own code. And like I said, different models excel at different things and they're worse at other things. So what we're seeing here is that depending if you're using Claude or Codex, they cluster in different ways. So for example, missing behavior, Claude more guilty.
¶17Semantic intent, Codex more guilty and so on. And by the way, if you want to see this entire blog post and all their findings, I'll drop a link down below. So if you're shipping code using AI, use Greptile to review that code. It really is great. I really do use it and so do teams like Nvidia, Mentlify, Posthog, Zapier, Substack and so many more.
¶18Ship like their teams. And there's also another layer to think about which is output speed. How fast are these tokens coming to you? Because there is a value in time. If it takes you 30 minutes to solve a task with GPT 5.6 and it takes you an hour to solve a task with Kimmy K3, then you are losing time by using Kimmy K3.
¶19So that's the first thing I want you to know. Not all tokens are created equally. And now you will know when a new model is released what to look for. So this is something that I do all the time on the hardest task that I give AI. And this also allows you to save a ton of money.
¶20I will have the best possible model do the planning. Then I will have a cheap and fast, but also really good model write the code, the actual execution. Then, I will have a different frontier model, the best of the best, do the review, the final review. And with these three things, I am saving money and getting the highest quality out of it. So, here's an example of what that might look like.
¶21So, I'm going to use, for example, Fable for the planning. I give it the hardest task I have, and I say, "Fable, come up with a full spec for this." It reviews my code base, thinks about how to actually build the feature that I'm asking for or solve the problem that I'm giving it. And the reason I give it to the best model is because that's the hardest part, is thinking from a high level, "How do I actually solve this problem?" Looking at the code base, trying to see around corners. That is what I give to Fable. Then, for the executor, I'm going to give it a cheap model.
¶22So, something like a Grok 4.5 or the Composer model from Cursor. These are two phenomenal, much cheaper models that are also really good at writing code. And it turns out you don't need frontier models to actually write the code. Once they have a plan, they can write the code quite easily. And so, you're paying a fraction of the price at this step.
¶23Then, finally, when all of the code is written, I'll give it to GPT-5.6. And so, that's another frontier model that looks again and says, "Okay, are there any bugs? Did they implement this correctly against the spec that was written by Fable?" And this is my workflow. This is what I highly recommend you do because the vast majority of tokens actually happens at the actual execution. Output tokens are much more expensive than input tokens, and there's a reason for that.
¶24The planning step doesn't use nearly as many tokens as the execution step. And for the planning step, it's actually using a lot more input tokens than it would during the execution step. During the actual code writing, the execution, it's using a lot more output tokens, which are much more expensive. Remember this, $5 per million input, $30 per million output. And then once again, during the reviewer stage, it's using more input tokens than it would during the actual code writing stage.
¶25So, this is the most expensive stage by far, and it's the one that's technically easiest for a non-frontier model to actually write. So, here's an example of what this entire flow looks like using Fable only, GPT-5.6 only, and then what I recommend, a mixture of models. So, for the planning stage, we're using a million input tokens because, of course, we have to read the whole code base, and 100,000 output tokens because all that needs to be output is a plan. Now, for the executor phase, we only have to read the plan, so far fewer tokens on the input side, and many more tokens on the output side because it actually has to write all of that code. And then finally, for the reviewer stage, once again, it's reversed.
¶26A lot of reading and just a little bit of output. And so, based on this math, what we find is Fable alone to accomplish this task is $81. GPT-5.6 coming in much less expensive, 46.50. But, when we use the mixture, we get the same results at a fraction of the price, 25.55. The clear winner.
¶27This is what we want to do. Plus, it gets the best results also. Not only that, there's one last benefit that I didn't even mention. Typically, these frontier models, the biggest, best models, tend to be slower. And so, by using a fast, efficient, cheap model, you are going to accomplish the task in less overall time.
¶28Now, the last thing I want to explain about tokens is actually a big deal. The future of how intelligence will be priced is being decided right now. There is this battle between closed source and open source, and a race between the US and China. But mainly, I want to talk about open source versus closed source. Right now, the biggest, best models on the planet are all coming from basically one of two AI labs in the United States.
¶29It's either OpenAI, or it's Anthropic. But they have the most expensive intelligence because each one of their tokens is the highest priced. Well, why does that actually matter? It turns out, these closed source frontier labs are making a ton of money on these tokens that they're selling. And that's their business model.
¶30They want to sell tokens, and they want to have huge margins on those tokens. And they're doing it. They're doing it quite well. But open source has really come into the conversation, especially in the last few weeks, as being a very competitive alternative to these frontier models. As I showed you, Kimmy is very very good.
¶31It does use twice as many tokens, but since everybody in the world can actually go and look at the Kimmy model, can actually download it, run it themselves, they can optimize it for their own use cases. They will figure out more ways to drive down the price of inference. All of the hyperscalers are going to be battling to give the lowest price possible on those Kimmy tokens, because they can all use it, and they can all serve it. And that benefits us, the actual consumers of those tokens. They're battling for the lowest price, we win.
¶32Now, on the other hand, the closed-source frontier models don't have as much incentive to bring their prices down. They're sitting there thinking, "Well, we have the best models on the planet. Only we and our partners can serve it, so take it or leave it." Now, of course, it's not quite like that. There is a duopoly currently between Anthropic and OpenAI, and they have to compete with each other. And we've seen the benefit of that simply by how many times the Codex team has reset the quota over the last few weeks in competition with Anthropic.
¶33Now, when open-source wins, the margins on the tokens drop substantially because everybody can serve it. All of the hyperscalers, anybody with a data center can serve these open-source models. And so, they're incentivized. They have competition, and so they need to drop their prices as much as possible. And so, what happens then is the profit in artificial intelligence goes away from serving tokens and goes into every other part of the AI stack.
¶34It goes to Nvidia and other chip providers because as the price of a token decreases, people use more tokens. More tokens means more chips. It goes to the hyperscalers and other data centers because as people use more tokens, they need to be able to serve those tokens. And so, they're serving a lot more tokens, and thus they're making more revenue. It even goes to the application level.
¶35If you're building software on top of artificial intelligence or using AI to build software, you're paying less, your software is getting better because you're using more tokens, and overall, it's just good for the application layer as well. And so, that's the battle happening right now. If the closed-source frontier labs win, then you most likely will continue to get the best tokens, but you're also going to be paying a premium for them. If open source wins, you will be paying less money per token. The closed source model providers like OpenAI and Anthropic will probably have a competitive disadvantage, and they will have to compete in other areas other than margins on the tokens they're selling.
¶36Now, I know this is a lot. I could probably make an entire video about the economics of tokens or tokenomics, but that's a little primer for you. Just know this stuff is happening right now. And if you want to stay up-to-date on the latest in AI, check out our newsletter Open source is so important. That's why I made an entire video about it, which you could check out right here.