New ChatGPT Model & Memory Features Explained (AI News You Can Use)

So this week in AI, something significant changed. It's not a new app, not a new model. Well, I guess it is a new model. But more importantly, it's how you're supposed to interact with ChatGPT's new main model. It shifted. Basically, instead of being super detailed, they now recommend you prompt for results. So what that exactly means and how the results differ from what you might have experienced last week. And also they claim that hallucinations were reduced by over 50%. I'd like to hear that. So what we're going to be looking at today among some other news stories in this week's episode of AI News You Can Use, the show where we look at everything that comes out and we pre-select the things that actually matter to your experience with these products as a non-technical consumer. Let's get into it. There's a new model inside of ChatGPT, but you wouldn't even see it at first sight because they simplified the interface even further. Instead of having the model selection up here, as last week, maybe we can pop it up on a screen. It's hidden in here and you can see GPT-5.5 Instant. This is the new model that I'm talking about and it's the default for, well, all plans. If you want to access some of the previous models, you can go to configure. And then in here, you have some previous versions. But really, for most people, it's just this. They're going to be using this model by default. And sure, it's quick, as the name suggests. Well, near instant. But mainly, they published a whole article talking about how you're supposed to prompt this thing differently than any previous models from OpenAI. I want to talk about that. Now, I want to be clear, this difference in how you're supposed to interact with the model doesn't just apply to GPT-5.5 Instant. Across all 5.5 models. But I think it's important, especially for some intermediate to advanced users who've been following this for years and have a specific way of doing things. 5.5 models work better with shorter outcome-first prompts, where you show what a result or what good looks like, rather than showing step 1, step 2, step 3, 4, 5, 6. If you have prompts that have the sequence inside of them, or you're used to saying, hey, go do this, then this, then this, especially maybe inside any automations or agents that you might have running or some prompts that you've been refining over time, it might be worth reevaluating right now if if a single goal works better with this model. So I'm going to show you an example, but the details of this guidance comes from OpenAI themselves in their developer documentation that was not very prominently featured, but I think is really relevant for everybody using ChatGPT now. So let's just try this. Memories off, no other context in here. Simple prompt. I have 5 video ideas. First, read them, then evaluate against my criteria, which is audience appeal, production effort, SEO potential, and channel fit. And then score them, sum the scores, rank them, and find the winner and explain the reasoning. "Ah, multi-step process. Here's the ideas below." Let's just run this in the Instant model. But this does apply to all 5.5 models, which, you know, these would be the other two. Now, here's the other way to prompt. It's arguably simpler. Look at that, way shorter. "Pick the 5 strongest of these 5 video ideas for my channel." And I add a little bit of context on the actual channel. And then I tell it what a good outcome looks like. "One clear winner with a 2-3 sentence rationale." That's it. Same ideas. Let's run it. And by the way, I'll run the same prompts on Extended Thinking to compare results. Let's start with the old way of prompting. So it ranks all of them, it creates this table, and basically decides that number 1 is the best. ChatGPT versus Claude versus Gemini in 2026. The reason for that is search intent, broadest audience, etc. When I compare it to the shorter prompt, it gives a different result. It says number 2, I built an AI agent that runs my inbox. Look, this is not a scientific study. This is me just testing things that I do with these regularly, and me getting a feel for it so you can get a feeling too. I personally, think that actually number 2 is the strongest one. Number 3, I know we asked for different things, we were more specific, but one, this is harder to write. But just this result that, you know, seems better because there's all of this reasoning. Well, honestly, this video would never do well. This is a very 2023 type of video for me and the channel here. And here's where it gets interesting. When I switched to the thinking model, it fought for 4 more seconds and actually arrived at the fact that no, result number 1 is actually the winner. AI agent running your inbox. In other words, when it had more time to think, it changed its opinion to what goal-based prompting did in just a few seconds. Isn't that interesting? Simpler prompt, more accurate result that I personally would agree to. And accurate is defined by me. Now out of curiosity, what did the simpler or shorter prompt, the outcome-oriented one get on Thinking Mode? It also got winner too. So again, this is not a scientific experiment. I'm just telling you that they publish that the prompting changed. From this example, I would agree it's probably not a cataclysmic change. Your stuff is still going to execute step by step. But especially in your everyday usage, it's just such a good habit to get into, you know, give it some info on you, give it what you want to do, and then tell it what good looks like. In my In my recent presentation at the AI Advantage Summit, I actually presented this framework. I called it the context sandwich. And it looks something like this slide right here. And then I also kind of jokingly noted that, you know, if you want to think of it as a hot dog, it might be easier to remember. And the same thing applies, just the visual changes. And what it really means is that, you know, the top bun is some identity on yourself, some context, then a task, and then what the result looks like, what good looks like, the goal, goal-based prompting. That's what works really well with this model. And I would challenge you to test some of your best prompts because I don't give this advice often, but it's easier to prompt for goals. Now it works better too. So that's a goal, but there's also the identity piece, right? The first part. Well, and that is usually managed in ChatGPT with the Memories feature, which they upgraded too. So let's have a look at that. If I go to personalization, there's this Memories option. You're probably familiar with this, but just a quick recap, it's basically a way for you to look into past chats and get tidbits of information that it stores in a vault that is linked to your account. And that vault of information basically informs new chats, meaning it remembers past interactions if ChatGPT decided to store the memories. Now this if factor was a big problem for me. I love to manage the context. Context is at the core of your experience with AI. So I want to have manual control over that. And previously memories were just a bit random. Now with this new GPT-5.5 update, they actually made updates to memories too. So check it out. If I say, tell me about myself, because memories are on, it can look into past chats, it can look into info on me. And then here under sources, I can actually see where it referenced a specific saved memory. From here, I could view all saved memories. I don't have many. I usually manage context myself. But with this updated. I don't know. I kind of like the fact that it showed me that it used info from this source, whereas up until now, I was really flying blind. And especially this ability to make a correction right now. So just click the three dots, say make a correction, and then I can update stuff as I go. Just, you know, not perfect. I still think manually crafting your context, especially on your identity, your goals, your values, is essential to getting more out of ChatGPT. That's what I preach across all trainings. That's where we always start. But this feature where it automatically gets it and you get some control, you get to go into sources, change it. I like it. It's the right direction. Good move. And hey, just real quick, if you're enjoying the show, make sure to subscribe. It really helps the channel. Now let's see what else is new. And then lastly, I want to point out this data point, which is, you know, kind of hard to prove. I'm not going to be able to like run a prompt here, run a prompt there and show you the results immediately. But they claim to have reduced hallucinations by over 50%. Just for context, some studies showed that hallucination rates went from about 20% to around 3% these days. Really depends on the model and type of questions. There's so much variability there. But just know that potentially we're really heading down and they're getting better and better at controlling, well, these confidently presented false facts. Facts. And this model is supposed to reflect that, especially when it comes to medical information, and also legal and financial topics. One fun thing about hallucinations, if you don't know them, usually they appear when you ask for something hyper-specific. So if you ask it for dates, for quotes, for numbers, and guess what? Medicine, law, finance, all very specific domains. I mean, doesn't get more specific than finance, right? A number is as specific as it gets. There's zero ambiguity. And when you ask it for something unambiguous, And it doesn't have it. Well, it's set up to be a helpful assistant. It gives it to you anyway. That's where a hallucination is often created. Just wants to help you even at the cost of presenting you with false information. Hope that helps. Okay, and then here are a few more points that we found in our testing with the team. One of them is that the search actually improved with this new model. Compared to its competitors, including Claude, the results had better structure. They're more concise, more to the point without dropping essential information. And then this old formatting of like paragraph, paragraph, paragraph with headings improved. Know how over the past year, it sort of feels like a lot of these AI models, they just give you like 15 paragraphs of different stuff, all of it sort of useful, but just a lot. That's how it feels to me. Well, now that improved a little bit. Even in our testing, when we looked for some of the best games of this year, it started adding these FAQs at the end of something. And I absolutely love that. I mean, that's sort of at least how my brain works. It's like, hey, give me the info and then clarify any questions that might be outstanding. Don't give me every possible angle to this. I like that. And I've never seen an FAQ like that as a part of an answer. So this is definitely something that changed just now. And look at how many images are just in this answer. It's really good, really concise. Honestly, a big reason why I preferred Claude lately is because it's just a bit more concise. I think a lot of that was due to their token shortage, which, by the way, fun fact, this is sort of just a Radar story, but they partnered with xAI and got a lot more compute and they doubled Claude code hourly limits, not the weekly limits yet from what I've seen. You can basically use more Claude code in a day now because of that partnership and they're going to be expanding their capacity. They were really, really limited by how little compute they had versus OpenAI that, well, has a ton from their various deals. Okay, back on track. Well, actually just realized that's the track. That's GPT-5.5 Instant. Less hallucinations. You can prompt it in an easier way for better results. And I like the result structure a lot actually. This is nice. One of those updates that seem minor, but in my everyday usage, it's gonna be huge. Oh, one last note from team testing when it comes to like websites, visuals, or games. Well, it's an instant model. It doesn't even get close to all the extended thinking models. So don't expect any big changes there. It's basically for these little everyday results where it does better with less. I have 3 more updates that I thought were really interesting this week. One is, well, on the same note from OpenAI, they released GPT Realtime 2. And it's basically voice agents with reasoning. It's kind of fun, no? So you know how a few weeks ago they did GPT Image 2, which was the first image generation with like extensive reasoning and web search behind it? Gather info, think about it, and then create an image that is actually accurate. Well, they're doing the same to voice agents now. And this is just in the API. It's not in the consumer-facing product ChatGPT yet. But they're saying it can reason at GPT-5 level. But it can live translate across 70 languages with stunning accuracy because it has this reasoning in the background. I'm going to try this over the coming week and fill you in. It just came out. And then also I want to point to the Microsoft Office related products from both OpenAI and Anthropic, because Claude launched Claude for Office into general availability now. So if you want Excel, PowerPoint, Word, and Outlook support from Claude, you can do that. Up until now, it was very limited. There was a research preview, higher tier accounts, many gating where you had to be on a waitlist. It was just messy and not available. Now it's out there. And OpenAI reacted with ChatGPT for Excel and Google Sheets coming out of beta. So, you know, both of them are now competing for this market. So if you work a lot in those, it's probably worth testing. We're already on it, seeing how these work on different prompts. If they work well, you can look out for a standalone video just on this. There you go. Those are the most significant stories from this week. If you want, here's a screenshot for you from our research database if you're curious about all the stories that we're kind of looking at this week. And yeah, that being said, my name is Igor, and I hope you have a wonderful week. AI news you can use, breaking down the things you care about. The kind of savvy you never want to lose, while Igor clears away every single doubt.