GPT 5.2 is out in n8n and its CRAZY (full breakdown + Demo)

GPT-5.2 just dropped and it's perhaps the best model we've seen so far. It beats models like Gemini 3 and Opus 4.5 in almost all benchmarks. So in this video, we're going to run through exactly what it does, all of the benchmarks, and we're even going to be using it inside of NAN. By the way, I just made a full prompt library specifically for GPT-5.2, so if you want to get access to that, I'll leave the link to my community in the description where you can grab it. But let's not waste any more time, let's get into the video. Okay, so first of all, we're just going to quickly look through the introduction to 5.2. You can see here on OpenAI's site, this is just introduction post to GPT-5.2, and they're saying GPT-5.2 sets a new state of the art across many benchmarks, including GDPVal, where outperforms industry professionals at well-specified knowledge work tasks spanning 44 occupations. And here it shows you the benchmarks between GPT-5.2 Thinking and GPT-5.1 Thinking. You can see here some of these are absolutely crazy. If we look at the knowledge work task, you can see GPT-5.2 Thinking is at 70.9% whereas GPT-5.1 was at 38.8. So you can see a massive increase there. And then also in software engineering, you can see an increase there. And then another software engineering benchmark, you can see also improved. The science questions improved by quite a bit. The science figure questions again improved by quite a bit. Competition maths, advanced mathematics, and then another advanced mathematics. You can see on all of these benchmarks it is improved by a lot. And then these two abstract reasoning benchmarks, you can see on this one is improved by quite a bit. But then on this one you can see it has improved massively. So that is something I've noticed. The reasoning and specifically the abstract reasoning has improved so much with this model compared to other models. Here you can also see some examples of what it can actually do. So if we look at the workflow planner, this is basically where it's creating spreadsheets for workforce-related tasks. This is what GPT-5.1 Thinking made, and then this is what GPT-5.2 Thinking made. So you can see, you probably can't see the actual numbers, but you can just see The layout and the design looks 100 times better. If you also look at a cap table, you can see 5.1 incorrectly calculated seed, Series A, and Series B liquidation preferences and left majority of those rows blank, leading to an incorrect final equity payout calculation. It also incorrectly inserted calculations in header rows. 5.2 completed all calculations correctly in an audible way. The design looks fairly similar, but like it says, 5.1 was wrong on a lot of these things. Whereas 5.2 was right on all of them. Then if we look at project management, you can see again the design, just like this one here, the design is much, much better with 5.2. It's made a full interactive dashboard here, whereas just the base 5 model didn't do that. It just created a simple spreadsheet, and you can see some of the information has even been missed. And then some of the more exciting stuff. If we look at coding, you can see it is dramatically increased on the software engineering benchmarks. And if we even scroll down here, you can see these are some examples of the things that 5.2 built. So here we have kind of like an ocean simulator where you can change the lighting. So if we increase the lighting, you see it increases and decreases. You can change the wave height, you can change the wind speed to make them go quicker. So very, very cool. You can pause it, you can reset it, you can go full screen. And then there's a holiday card builder. This one's really, really cool. Obviously a little bit more simple, but you can see just the animations, the little noises. I'll be quiet so you can hear the noises. Very, very cool. It's made these noises, it's just natively inside of the thing that is coded here. You can create little holiday cards, so you can type in things here. You can also do animations like shake, uh, you can change the snow type, you can change the sound to be on or off, you can delete these little emojis, you can add them in from here and they make those little noises. You can make them bigger or smaller, you can rotate them. So very, very cool. And then there's a type brain game. So we're actually going to try this out quickly. This is where you can, um, type things out and it's basically like a game. So if we go harbor avenue, uh, bridge. Okay, very cool. I'm not gonna, um, sit here and make you watch me do a typing game, but you get the idea. Very, very cool, very complex things that it's coding. GPT-5.2 represents the biggest leap for GPT models in agentic coding since GPT-5, and this is a SOTA coding model in its price range. The version bump undersells the jump in intelligence. We're excited to make it the default across Windsurf and several core dev workloads. And that is from the CEO of Windsurf. Also want to show you the benchmarks. Apologies for the low resolution image here, but as you can see, if we just take, uh, 5.2 Think compared to Opus 4.5 from Claude and then Gemini 3 Pro, you can see software engineering, it is the highest out of all of these. It's actually much higher than Gemini 3. And then the science questions here, you can see it is again much higher than all the other ones. Gemini 3 is a little bit closer on this one, but we're still closer with 5.2. Scientific figure questions— this is the only one where Gemini 3 slightly comes out on top. You can see by 0.3%. And then Frontier math, you can see it just— 5.2 thinking is better on this first one here, but then Gemini 3 is a little bit better on this second one. And then competition math, uh, 5.2 is way better than the other ones. Abstract reasoning again, it's a lot better than the other ones, especially on this second abstract reasoning benchmark. And then knowledge work task, it again is by far the best out of all of them. So you can see compared to these other models, Gemini takes the cake on I think 2 benchmarks there by just a hair, but all the other ones, 5.2 Thinking is far better. So if we jump into an actual demo here so you can see it in real life, essentially what I've done or what I've built here inside of an 8N is just a system that can take data from multiple different sources like Slack, your CRM, your emails, and put that all together. So just bunch up into one long list of data, and then we're going to pass that to 5.2 and get it to generate us a daily report. So essentially the kind of things we're testing out here is 5.2's capability to handle a lot of different data, and then also its reasoning capabilities to kind of find data from different sources and piece them together and then come up with a full report. So we're going to be testing its reasoning abilities and also its context window abilities. So I'm quickly going to show you the different data sources that we have here. I didn't connect these up to actual data sources, I've just put demo data in Google Sheets. So I want to show you through each of them. First of all, we have demo CRM data. We have contacts, deals, and companies. And in here for contacts, we have things like contact name, contact email, they all have a record ID, the company name, and then the deal value is empty because it's there's not a deal, the description, the urgency, and then the last updated date. And then we have the same thing for deals with the deal value as well, and then the same thing for company. And then we have the demo emails. So these are just going to be unread emails that loosely relate to the CRM. So we're going to pretend this is all for the same company. So there's going to be some, uh, emails in here. So like Alice@brightflow.io, you can see, is also in here. So we got an email from her Um, and that said the subject was Proposal Next Steps and the snippet was, hey, quick one, can you confirm the implementation timeline and who would own onboarding on your side? So we've got a bunch of unread emails like that. And then we have Demo Slack Messages. It's just going to be from internal team members, as you can see. So just lots of things relating to deals and companies and proposals. And then we also have Operative Contexts, and this is just going to be things like pause all outbound communication until January 5th. Due to end-of-year capacity limits, and then deals over 15,000 required Dan's approval, that kind of thing. It's like this variable information. If you're running a new offer, if it's around Christmas time, or you've got a new team member that's handling different stuff, it would go into operative context. So all of that information is going to be taken in NaN. So we're just going to return all the rows from all of those sheets, and then we're going to aggregate them together so they just put into one item, and then we'll merge all of that information together so it's in one data set. Then we'll simply just aggregate it so we make sure it's one item, pass that to our AI agent here, which is of course 5.2, and then just email that to ourselves. So this bit should all be pretty self-explanatory. It just takes the data from those Google Sheets. I do want to show you inside the prompt for 5.2. So as you can see, we've linked up 5.2. We haven't done 5.2 Thinking. There are a few different, uh, 5.2 options. There's the 2025, the 11th of December one here, there's the Chat Latest, there's the Pro. And then there's the Pro with the 11th of December again. So we've just gone with the default one, which is GPT-5.2. And then we've essentially just given it all of that data that I just showed you. And then this is a prompt that we have in here. So role, you are an autonomous operations manager AI. You act as a senior internal operator responsible for prioritizing work, enforcing operational rules, deciding what requires human attention, and suppressing noise that does not warrant action. You do not behave like a chatbot assistant or summarizer. You behave like an experienced operations lead making judgment calls under constraint. And then we have the task in here, so basically what it's actually doing. So we just tell it the data it's going to receive and then what it's actually doing with that data, with a few guidelines in here. For context, you are given 4 data sources representing the same company: CRM updates, things like contacts, deals, and companies, deal value lifecycle changes, intent signals, and then unread emails, and then internal Slack messages, and then operative context. Always read and apply operative context first. Never violate a high priority rule. As you can see, kind of like rules on how it should answer and what it should be doing. And then output format, return a daily ops brief in the following structure. So it should look like this with escalations for human action required. So for each, give the item, the summary, why it matters, why it cannot be automated, and then relevant rules considered. And then the automated actions. So the automated actions, things that can just be automated instantly. And then the logged ones are kind of just no action for these ones, don't really need to do anything. And then notable observations, things like patterns, risks, emerging opportunities. And then we've just told it to output in clean HTML. HTML. And the reason we get to do it in HTML is because when you send stuff over in Gmail, you can send it in HTML and it'll automatically render. So it just means we can have it looking a little bit nice and structured with, you know, bold text and things like that. So we're going to run this through and then we'll look at the results of it. So you can see here it's just grabbing all the information from Slack, CRM, the emails, and then the operative context, aggregates all of that, merges it, and we get 4 items. So we just aggregate that into one long list that will look something like this. If I show you, so we get data 0, data 1, data 2, and data 3. And then inside of all of these, data 0 is going to be the Slack information. We can look here, these are all the Slack messages. And then data 1 is going to be the CRM information. And then data 2 is going to be the email information. And then data 3 is going to be the operative context. Pass that all to this agent here. So it looks something like this. This is the data that we passed to it. It's just, you can see there's so much data here. In just one big JSON body. If we actually expand this, you can see how much data there is here. So much data that has to go through. And not only does it actually have to go through this and be able to ingest it, it has to get insights and go through that output structure that I showed you. So there's quite a lot of work for it to do. And this is what it gave us. It sent us over an email, so we'll have a look. Okay, cool. So it says daily ops brief, 2025, the 13th of December. So escalations, the deal, uh, this is the deal ID obviously with Fanova for £25,000. This is the summary. Finova requested timeline plus security review. From this email, an internal Slack flagged a security review risk. Deal was already noted as pushed to next quarter in CRM, but inbound request still needs correct ownership handling. Why it matters: high-value deal with compliance gate. Uh, why it cannot be automated: security review response requires a designated human. Relevant rules considered: all security reviews for Finova must be handled by Tom. Any deal above £20,000 must be summarized in the daily ops brief. Pause all outbound until January 5th, and escalations must include rationale why automation is insufficient. So this is really, really cool. As you can see here, it's grabbed this, it's understand that it's over £20,000, so it must kind of flag it up and do a little section on it here. And it says they've requested a timeline plus security review. So it's grabbed that from an email. Obviously we've got an email from them asking for this, and then also saw in Slack that they flagged a security review risk. So this is the Slack idea of that message. So from there it's gone into the CRM, found the information on this deal, And then it's found also the email containing information about this and has also found the Slack message containing information on this. And then also some information on the company. So it's the same company, Finova. This was about the deal. This is about the actual company. Finova Finance asked for a VAT number plus invoice address plus payment terms from this email. Slack echoed, can someone send? So again, very, very cool. It's found the email from them and basically said they need this doing. And it also found in Slack someone recognized this and said, can someone send this over? So it's just flagging up for us. Automated actions: route inbound Fonova security review email to correct owner, route inbound Fonova finance email to Laura, no external response, enforce pause outbound guardrails across queue. So you can see here, this is really, really cool. It has, um, basically said that this task as a whole needs to be done by a human, but what can be automated is routing the inbound Fonova finance email to Laura. So not the actual sending over the information to the company themselves, but just sending that email over to Laura because it knows from the operative context that this is something that's dealt with by Laura. So very, very cool that it just automatically knows that. And then all this stuff that's logged with no action, you can see it's just picked up on the deals, the email. Uh, for example, North Peak had some pricing questions. Any reply would constitute outbound communication, also violate intentional pause. So it's seen that we're not doing any outbound messages until January 5th, and it's picked up on that. So it's basically said don't take any action on this. And then notable observations: rule conflict risk, multiple inbound Email requests, responses ASAP. But OP001, this is just going to be from operative context, forbids, forbids outbound communication until January 5th. Current best practice is internal routing plus preparation, plus preparation. No external replies unless leadership explicitly authorizes exceptions. And then also, Fanova is the highest impact thread. So basically just focus on this Fanova deal and the company, uh, tasks that need to be done. Brightflow growth signal, headcount up 30%. Suggests expansion potential, but response work is frozen for January 5th with tight implementation timeline arts prepared. So very, very cool. And then the final check: no external outbound actions proposed during the operative context 0013. So that's just until January 5th. All escalations are internal ownership escalations that can be actioned now and do not violate no outbound. So it's basically saying everything that it wants us to do as a human that can be automated or needs to be done as a human It's just internal escalations, and it tells us exactly how to do that and exactly why we should be doing all of that. So very, very cool. I mean, of course I haven't tested all of the capabilities of 5.2. I still want to test out the coding capabilities, but I've seen that is very, very good. I think this model is absolutely amazing. It now being the default inside of ChatGPT is great. Like I said at the beginning of the video, if you want to grab the prompt library for 5.2 so you can get started with it straight away doing some really, really powerful stuff, then I'll leave the link to my community in the description where you can grab that.