GPT 5.6 explained in 8min..

It's not like OpenAI never released the same model in different sizes before. I'm sure we've all been confused by their naming convention in their models. I mean, just look at all these models. How do you expect anyone to choose the right model for the right job? Now, OpenAI's newest release of GPT 5.6 model consolidates into three. The GPT 5.6 model was announced as Saul, Terra, and Luna. Saul being the biggest and the most advanced and Luna being the smallest. And if you've been using anthropic models before, it mirrors a lot like their haiku sonnet and opus variants except until recently of course we had a new mythus tier which later nerfed into yet another model called fable. I know a lot to keep track of as a consumer but it actually gets worse. These models are just from top two frontier companies here in the US. And when we start to look at open models from China, China has long been contributing to the collapse in pricing by undercutting the token prices. And their models are getting more and more efficient while catching up to the US, notably from Deepseek, Quen, Kimmy, and GLM. What's even worse is that labs like OpenAI and Anthropic that are supposed to be leading the pact are also butdding heads with the US government that's trying to regulate highly capable models like the GPT 5.6. six from OpenAI and Methus from Enthropic. So in other words, there's an upward pressure as Chinese models continue to catch up in capabilities while also offering them at the cheaper price. But there's also a downward pressure from the government trying to regulate and slow down market adoption of highly capable models which makes it really tough for Frontier Labs like OpenAI and Anthropic. And from the consumer's perspective, while we are seeing tokens being offered at a cheaper price, tokens are increasingly becoming privatized where more intelligent tokens are directly to select few companies first that are part of the regulatory body. And the growing fear is that eventually the thesis on why Chinese labs across the board release their models for free as open-source might come to a stop if China ever matches the frontier capabilities. This could easily paint the worst case scenario from the end users. Now let's briefly look at pricing. When we look at pricing for GBT 5.6 model in the context of the five layer cake that Jensen Huang often uses, we see that the model is offered in three different pricing. Soul for enhanced and extended reasoning at $5 per million input tokens and $30 per output tokens. Terra for a more affordable but capable model at $2.50 for input and $15 for output. And finally, the Luna model at $1 input and $6 output. To put it more simply, you're looking at a 2.5 times price jump from Luna to Terra and a five times pricing jump going from Luna to Soul. So naturally, given this, we will likely see the demand profile from the application side consolidate into Soul being used for more advanced tasks like coding and research, Terra being used for more day-to-day agentic use cases, and finally Luna for more fast-paced sub Asian tasks. Now, one interesting part to look at when it comes to what OpenAI announced is actually in the chip layer. OpenAI hinted that their GPD 5.6 model, specifically the sole model, could achieve up to 750 tokens per second when paired with a Cerebra's chip. And this kind of promise isn't something that Chinese labs can match at scale because of the systematic advantage that the US has over China at the infrastructure and the chip layer. In other words, just because you have a really capable model at the model layer doesn't necessarily mean you can achieve fast inference for people to use at the application layer unless you have the infrastructure and the chip that can support them. And according to OpenAI, they're hinting that starting July, the sole model will be offered also on the Cerebrus chip as well. So what this means is that GPT 5.6 six announcement is not just news at the model layer but also news at the infrastructure and the chip layer as well as they leverage the inference ecosystem. So while China can still compete at the model layer, this kind of inference scaling isn't something that they can realistically deliver for a long time now. We can't really talk about GPT 5.6 model announcement without talking about the growing involvement of the government regulation. But first, a quick word from Immersion AI sponsoring this video. If you're like me who just wants to build and build, you're always looking for a good platform to build your next application. Emergent is an agentic AI app builder. And here I'm adding in my prompt in the search bar describing my requirements. I'm creating this app for my wife who's an esthetician and she's been wanting a booking and scheduling software for her appointments with her clients. An emergent asks questions to clarify my requirements. And as you can see, it goes ahead and builds the entire app, the database, the backend, login and deployment. Now, instead of creating just a mockup, Immer takes the entire scope of the web development and executes them like having a small AI dev team working in parallel. This means emergent handles the entire setup of the repository, choosing the right stack for the job and all the meticulous details that typically goes into developing applications. Here I'm clicking around to see what it generated, which is pretty close to what I had in mind. And once I'm satisfied with the product, I can just publish it with a single click, and a URL is now generated, which means people can now access this application. What's great is that you own the code emergent generated and you can pull directly into your local editor to customize without lockins. And if you're a builder or just need a product that can help get things done, Immersion can help you build and ship it. Link in the description below. Now, when we look at the fact that GBD 5.6 six model is on hold until the government essentially green lights for public release. It's hard not to look back and see what happened with Anthropic with their release of the Mythis 5 model which was a total train wreck to say the least. But even though OpenAI reached the same predicament for now as Anthropic, the nature of the release for OpenAI is actually very different than Anthropic. And here's why. When we zoom into the month of June, for example, the US government first signal through an executive order to help regulate how powerful these AI models are becoming. And this order was not really a strongly worded mandate that demanded labs like OpenAI and Enthropic 2 comply for permission before releasing, but it was more like a voluntary framework where they can opt in to get government input. But as we all know and as we all seen later in that month, Enthropic released the Mythus 5 and the Fable 5 model and Enthropic has been marketing these models to the public as a high-risk and dangerous model which was later ordered by the government to shut it down at least in the practical sense by mandating anthropic to restrict all foreign nationals from using what they implied as a highly dangerous model. Now in OpenAI's case things are a little bit different. Now, even though OpenAI did release a model that is at least on par, if not better than Mythus 5 and Fable 5, when we look at the benchmark, at least OpenAI was able to see Enthropic as precedent in how this executive order actually looks like in practice. After seeing how Anthropic had to essentially rugpole their entire users from accessing their most highly anticipated models, OpenAI opted in to take the voluntary way by asking the government ahead of time before releasing them publicly. And even though the result was not that different, OpenAI chose to be in compliance with the order and at the same time stating that OpenAI doesn't believe that this process should be long-term since it takes away power from users and that this is just a short-term step as they continue to work with the government to iron this out as more and more advanced models get announced.