¶1Open AI just dropped their prices by 80% and that's awesome, but it's also not even the most impressive thing. They used their best model, GPT 5.6 Soul, to improve itself to find efficiency gains. That's how they were able to drop the price so much. I have so many thoughts about this because this is essentially recursive self-improvement and there are so many implications. Let's go through them.
¶2First, let me show you the news. So, here's Sam Altman, major price cuts today, an 80% drop for GPT 5.6 Luna. Luna is the smallest of the three models in the GPT 5.6 family and it's actually really good. And when I show you the charts comparing it to other models including Fable, you're going to be even more excited about this price drop. So, now it's 20 cents per million input tokens and a dollar 20 per million output tokens, which is incredibly inexpensive.
¶3That is cheaper than the frontier open-source models coming out of China. And I know I've been talking a lot about this before. It is not strictly about the price per token. What is more important and really only important is the cost per task completed. And the reason is is because for example, Kimmy K3 is half the price of 5.6 Soul, but it takes twice as many tokens to accomplish that same task.
¶4Thus, they're effectively the same price. Now, for GPT 5.6 Terra, which is their middle model between Luna and Soul, there is a 20% price drop. Now, it is $2 per million input, $12 per million output. Now, GPT 5.6 Soul, the big model, the model that was able to achieve some of these performance gains, did not get a price drop. Well, not exactly.
¶5So, what we did get is a fast mode in the API. You get two and a half times the speed for two x the price, same intelligence. Well, that sounds nice, but it doesn't sound like a price decrease. Well, it kind of is. Previously, the fast mode options were one and a half times faster for two times the cost.
¶6Now, we get two and a half times faster for two times the cost. Same cost increase for fast mode, but much faster. That is effectively a price decrease. And so, let's look at this. This is the artificial analysis intelligence index, and what we're seeing on the Y axis over here is the overall intelligence index score, and importantly, and I'm so glad that Sam Altman shared this metric, cost per task on the X axis.
¶7Now, look at this. What you want to see on this chart is as high up as possible and to the left as possible. And what do we see here? GPT 5.6 Luna sitting right there. Now, each of these dots represents different thinking efforts.
¶8This top one is the max. So, GPT 5.6 Luna max. Now, what we see is it is slightly ahead in terms of overall intelligence to GLM 5.2 max, which is an open-source model out of China. And because it's all the way over here, it is a fraction of the price. So, GLM 5.2 max looks like it's coming in around 25 or 28 cents per completed task, whereas GPT 5.6 Luna is coming in at about 6 cents.
¶9Again, a fraction of the price, and it's slightly ahead. Here's Claude Opus 5 low, coming in slightly lower than GLM 5.2, slightly lower than 5.6 Luna, but it is really expensive at about 40 cents per task completion. So, Luna is just going to be an absolute workhorse of a model. And all of that is great. But, that's not the crazy part.
¶10The crazy part is they used their Frontier model GPT-5.6 Soul to find all of these performance gains. That allowed them to drop the price of their other models. So, just yesterday after deployment, we applied GPT-5.6 Soul to advance the frontier of efficiency by making itself more efficient to run. The results: 20% lower serving costs from production GPU kernel improvements, 15% better token generation efficiency from improved speculative decoding. And if you don't know what that means, it doesn't really matter.
¶11It just means that they were able to improve the overall efficiency of the model. And they actually wrote up an entire blog post about the technical details, the specific performance improvements that were discovered by GPT-5.6 Soul. I'm not going to go too deep into it, but I want to read a few of these quotes because it is stunning. Throughout the entire article, they're saying, "Yeah, 5.6 discovered it." Check this out. So, 5.6 Soul and Codex helps us analyze production traffic, identify previously overlooked sources of imbalance, test new routing strategies, and constantly tune these heuristics.
¶12And so, just imagine this. Every time you submit a prompt to ChatGPT, and then imagine that multiplied by the millions of people that do that per day, GPT-5.6 Soul with Codex is just constantly looking at the production data, constantly looking at your prompts, constantly looking at where the inefficiencies are, and coming up with potential improvements, testing those improvements, and iterating over and over again. And if that sounds familiar, that is also known as a loop. The goal of the loop being to improve the efficiency of the model. And so this really shows the power of loops, which I've been talking a lot about lately.
¶13In fact, I made a loopy skill that you can install that helps you create and discover loops for your prompts. I published the skill, and you can try it out in the sponsor of today's video, Hyper Agent. I'm going to drop a link down below to my published skill, so you can get it directly in Hyper Agent. Hyper Agent is awesome because it is like a team of always-on agents working for you. And you can use models from any provider, including the model we're talking about today, which is GPT-5.6 Luna.
¶14You can use Kimmy K3. You can also use GLM-5.2 Fast. All the models are available very easy for you to try out. Each agent gets its own cloud environment with its own login, completely connected to the tools that you already use. This is much more than a chatbot.
¶15And of course, an agent can delegate to sub agents, very organized and very cost-efficient. And with the marketplace, it's basically like an App Store, but for agents. And of course, you can find my loopy skill in the marketplace for free. So go find my skill on the marketplace, link down below. You get $500 in credits to try it out.
¶16So thanks again to Hyper Agent. Now back to the video. We also used Soul to optimize the model's forward pass, the computation that transforms inputs into next token predictions. They also found work that can be precomputed, avoided, or parallelized. And it even rewrote and optimized our production kernels.
¶17So just imagine having all the power of this frontier model running 24 hours a day, highly parallelized, looking for optimizations at all times. This is the power of loops, but this is also the power of recursive self-improvement. And this also makes me a little bit nervous. This is the point at which I look and I say, well, if the top two leading AI labs, OpenAI and Anthropic, already have such a big lead and they have momentum and we are at the beginning of recursive self-improvement, how does anybody else catch up? And that's a question I'm trying to navigate every day.
¶18Now, listen to this. It improved its own draft model by designing and running hundreds of experiments on its architecture, testing changes in size, structure, and features. Now, does that sound familiar? That sure sounds like Karpathy's auto research project. A few months ago, Andrej Karpathy, one of the most well-known AI researchers in all of the industry and also now an Anthropic employee, published this open-source repo in which he basically described what is a loop and the objective is improve the efficiency of training a brand new LLM from scratch.
¶19And the crazy thing is, the model came up with its own experiments, ran the experiments, reviewed and analyzed the output of the experiment, and then iterated over and over again. And it was actually able to find real improvements that Karpathy himself could not find. Now, he was applying this to a small language model. Imagine you have the compute of OpenAI and imagine you have one of the best models on the planet and this loop, this auto research loop, gets run 24 hours a day. This is truly an automated AI researcher.
¶20Now, back to this chart, you need to be using Luna. Luna was already a fantastic model and now you can run it on max mode and it is so inexpensive. This is a major, major upgrade to all of our lives. Now, the one thing we didn't get is a reduction in price for Soul. And I was thinking about why.
¶21They almost certainly were able to eke out performance gains, efficiency gains, for their biggest model Soul. But why didn't they reduce the price? They were able to drop the price of Luna substantially. They dropped the price of Terra by 20% but nothing for Soul. Well, here's my thinking.
¶22I think they actually were able to find major efficiency gains for Soul. But Soul is their cash cow. That is where the majority of revenue comes from. It is not from the workhorse models. Although the total volume of tokens might favor Luna and Terra, although I'm not sure, almost certainly the vast majority of revenue comes from the best models.
¶23Not the workhorse models. And so with those efficiency gains that they probably got for Soul, they're just able to increase their margins that much more. Now, what may have caused them to focus so much on efficiency and dropping the price? Well, it's probably this, Kimmy K3. The Chinese open-source model that was just released that is extremely good, near the frontier, but it's also much less expensive.
¶24Now, keep in mind Kimmy K3 generally takes twice as many tokens to accomplish the same task and it's half the price. So, it's effectively the same price, but a lot of people look at the price per token and think, "Wow, that is cheaper." But the other thing is as more inference providers get their hands on Kimmy K3 and more researchers try to figure out how to get more efficiency out of it, more people will be looking at Kimmy K3 and other open-source models and the rate at which they will discover new efficiency gains will be higher than what a small closed team can do. And so I think OpenAI's probably just trying to get ahead of that. But here's how crazy this price reduction is. Look at this.
¶25This is the cost per intelligence index task on artificial analysis. We have all the way at the top at $2.75, Claude Fable 5. Then, here's GPT 5.6 Soul Max at $1.54, much less expensive. Kimmy K3 coming in at 72 cents. And now, here's GPT 5.6 Luna Max at 6 cents.
¶266 cents as compared to all the way up here, $2.75. Now, obviously, Fable 5 is much more capable, but the price difference is astounding. And you can really see that in this graph. Here's GPT 5.6 Luna Max coming in at 6 cents. Here's GLM 5.2 coming in at 26 cents.
¶27Grok 4.5 coming in at 50 cents. And these are all basically the same intelligence. And you can just see, 5.6 Luna is in a land of its own. And this green quadrant, top and to the left, is the best. So, again, comparable to basically the the same intelligence as Claude Sonnet 5 Max, but a fraction of the price.
¶28Claude Sonnet 5 Max is $1.80 per task completed versus 6 cents. Now, I want to speculate for a little bit. What I actually think might be happening is the two major closed-source frontier labs, OpenAI and Anthropic, this is their playbook going forward. They train these massive models, really inefficient, massive, expensive-to-serve models, and then from that, they use it to bake almost as good, but far more efficient and far less expensive models for the rest of us to use. And yes, I said for the rest of us, because I actually think there's a pretty good chance, especially for Anthropic, that they do not release their absolute frontier models like Fable publicly anymore.
¶29Or at least wait a very long time. We already know Anthropic had fable all the way back in January, months before they released it publicly. And so I think this is actually a pretty strong strategy for these AI labs. Train the biggest best model you can, then use that to train smaller versions of it, highly capable, but much less expensive, make a ton of revenue for that, and then use that revenue to train your next big model. And at the same time, that biggest model is training the next biggest model, and nobody else has access to it.
¶30This might be recursive self-improvement. This might mean that Anthropic and OpenAI cannot be caught. They already have too much momentum. They already have the best models training the next best model. They have all the compute, and they have all the ingredients in the world.
¶31And that's why I'm such a big proponent of open source. I think open source is the only thing putting competitive pressure on these two labs. And one last thing, having such a substantial price drop might change the math on which provider you go with, Anthropic or OpenAI. Who are you going to pay your subscription to? I actually made an entire video about this exact topic, so it can help you decide.
¶32Check that video out right here.