Build ANYTHING with Gemini 3 Pro and n8n AI Agents

Google just dropped Gemini 3, and there are some big claims about this model already. On the official release blog, it says a new era of intelligence, and that Gemini 3 is our most intelligent model that helps you bring any idea to life. So you can go ahead and read through this, and you can get some information about the model, the DeepThink, and some of these benchmarks, but I'm going to be going over a lot of this stuff with you guys today. So I've been seeing some crazy demos all morning where people have been coming over here to the Google AI Studio and using Gemini 3 Pro Preview for completely free. Right now it's only actually over the API that you have to pay for Gemini 3 Preview, and that is where you get this input and output cost per million tokens. But like I said, I've played around with it a bit in here and I've seen some really crazy demos of people actually building things because what you can do is you can describe your idea right in here. It almost looks like a lovable or a base 44 type of interface. And so once you actually fill in a prompt of what you want to build, it can do tons of crazy things. You can make games, apps, landing pages, all that sort of stuff. And so if you just look at some of these examples that have been built with Gemini 3 Pro, they look absolutely incredible. Gemini 3 is really, really good with like the whole UI design, but also with all of the code that actually goes into the backend of all of these different landing pages. And you can also see those ideas can be turned into games. So right here we've got a runner game, right here we've got some sort of pilot game. You've got all these different interactive examples of games that you can build just because of how good Gemini is at reasoning and coding. So I'm sure all over X and LinkedIn and YouTube, you're gonna see some crazy examples of what people are able to build with Gemini 3 Pro. But today what we're focused on is actually using it over API and how we can actually integrate Gemini 3 Pro with n8n to power our AI automations and make our AI agents faster, smarter, everything like that. So I've done some playing around with it myself, but I wanted to do a live experiment with you guys where we're connecting to Gemini 3 Pro in n8n. We're then gonna test its ability to understand images. We're gonna run a massive amount of context through it and ask it some pretty tough questions and see how it can handle them. And we're going to have Gemini itself actually build NNN workflows for us and see how well it does. And throughout these 3 experiments, we're actually going to be comparing Gemini's results to different chat models so we can see how much better it really is for the price. So before we actually hop into NNN and start experimenting, I just wanted to real quick go over some release info as well as those benchmarks that we saw on that release blog. So today is November 18th and Google just dropped the model today. Right now it is being called Gemini 3 Pro Preview. So this is available in the Google AI Studio like we just saw, and it's also available through developer platforms. So if you go to Gemini API documentation, you can see we have a Gemini 3 developer guide, and it will walk through what it is, how you can use it. It also does some pricing and context window information and some API features about Gemini 3. And I wanted to real quick zoom out and show you guys a comparison between Gemini 3 and then sort of the 2.5 family where we had Flashlight, Flash, and Pro. Now you can see they roughly all have the same in and out context window, meaning they can take in about a million tokens and output 64,000 tokens. But the pricing is where things get interesting. So starting with Flashlight, we've got $0.10 and $0.40, and this is across a million input or output tokens. For 2.5 Flash, we've got $0.30 and $2.50. And then for 2.5 Pro and 3 Pro, we have, you know, it's a little more expensive, but it's also adjusted if the tokens go over 200,000. But what you can see is that Gemini 3 is going to be more expensive than Gemini 2.5 Pro. But I think that's going to be justified because it's going to be the most advanced, the best reasoning, and it's truly multimodal. So here are those benchmark statistics that we saw on that release blog, and I'm just going to dive into a few of them here. But we're comparing Gemini 3 Pro, Gemini 2.5 Pro, Claude Sonnet 4.5, and GPT-5.1. So what we're comparing here are kind of the best and latest top-tier closed-source chat models. And so for every benchmark, whatever's in bold is the one that won across that category. And if you look through all of these, I mean, Gemini 3 Pro is basically leading every single category. Some just by a little bit, but some like the Screen Spot Pro is leading by pretty much double the next one, which was Claude Sonnet 4.5. And this just means the chat model's ability to look at something like an image and understand it. And then this is just that second half bottom section of this, the same image, the same benchmark statistics. And we've got some different ones to look at here. And honestly, one of 'em that I think is the most interesting interesting is this Vending Bench 2 benchmark, which if you guys haven't heard of this, it's actually kind of a cool experiment they do. They have these AI models basically run a virtual vending machine business where the AI model would be in charge of ordering more inventory, adjusting prices, things like that. And the idea is that AI right now is typically really, really good at sort of like one-off tasks and short-term objectives. But when you start to give it a long-term goal that it needs to sort of accomplish and iterate upon over months, that's where they start to fall off a bit. But seeing that the leader before this was Claude-Saunette 4.5 with about $3.9K net worth on their vending machine, you can see Gemini 3 Pro came in with almost $5.5K. So it's really cool to see that these AI models are moving more towards being able to actually take on long horizon agentic tasks. So anyways, that's enough of these boring slides. Let's hop into NN and just start playing around with the model and seeing how it actually works. All right, so I told you guys we had some different experiments that I have prepared, but first what I want to do is just show you guys if you've never connected to Gemini before, how do you do it in n8n? And there's a few different ways. So the first way is you go ahead and just type in Gemini and you can pull up the Google Gemini node itself. And now what's cool about this is you have different types of actions. You can analyze audio, you can analyze a document, you can upload files, you can analyze images, you can analyze videos or generate videos, and then also just send a message to the model. And so if you choose one of these operations, what it's gonna ask you to do is create a credential where you would basically need to go get a Gemini API key. And so if you go to the Gemini API documentation here and you could just type in like Google API, whatever, you can see up in the top section right here, you have a button called Get API Key. And this will take you to Google AI Studio where you can go ahead and create a key and just put in your billing information right here. And so what that's going to do is it's going to give you an API key that looks like this, and you would just need to copy that and then go ahead and throw that right here into Anyttyn. And that's really cool because if you look at a different node, like for example, an OpenAI node, it doesn't have the ability to analyze a document. So this one has more power, but you can also take the chat model approach. So let's say we're throwing in an AI agent, and when you have an AI agent, of course you have to give it an AI model to use as a brain. And so when I click on chat model, you could go ahead and grab Google Gemini Chat, and this is actually gonna be the exact same API key and credential that you already set up here. So if you did this there, you'll have this all configured already as well. But if you're like me and you'd rather use OpenRouter, you can certainly do that as well. And OpenRouter is basically a connection that lets you route to tons of different chat models. So this is what it looks like when you go to the page openrouter.ai and you click on models. And you can see we've got Google Gemini 3 Pro, but we also have tons and tons of other types of models. And so the reason I like this is because I can just keep all of my billing information in one spot rather than having a Google key, an OpenAI key, an Anthropic key, all these other keys. So once again, you'd go ahead and sign up and then you'd be able to get an API key for OpenRouter, add your billing information. And then once you have that connected, you could go ahead and search for Gemini, 3 Pro, which is right here, Gemini 3 Pro Preview, and then you'd be connected in NNN. So fundamentally what's going on with all 3 of these methods is we're using Gemini 3 over API because we're talking to it from NNN. And so when you're using a model over API, there are a few settings that you can look at. So before I show you guys what I mean by that, if we go back into the Gemini 3 API Developer Guide, you can see that we have new API features in Gemini 3. The first one is thinking level. We also have media resolution, we have temperature, we have thought signatures, things like this. So the reason that I wanted to bring this up is because if you want to look at tweaking the model behavior a little bit, you have to adjust the settings. And thinking level might be one of those settings that you want to adjust. And what you'll notice here is that right now we have low, which minimizes latency and cost. And sometimes you might want to use Gemini 3, but not have to pay for all that reasoning and time. And high is the default setting. It looks like Medium is coming soon, but right now it's not supported. But if you go in here and you've got an AI agent, and let's say you've connected it to Gemini 3 Pro here and you want to make the thinking level low, you would probably think you could go here in the settings and change that, but that's not there. What we have is max number of tokens. We have temperature, top K and top P, and we have safety settings. And then you think, okay, well maybe if I connect to Gemini 3 Pro and OpenRouter, I'll be able to change that. Well, in OpenRouter, if we're connected to Gemini 3 Pro, we have a few different options and a few extras. But we still don't yet have the thinking level. And I'm sure what's gonna happen is it will come to this node over here, but right now it's not there. And then you might think, okay, well, let's just try the native LLM call here to Gemini 3 Pro. We're gonna choose the right model and add options down here. We have a few other ones now. We have a thinking budget, but that is not explicitly the same thing as setting a thinking level. So the way that you would want to actually be able to control that and know for a fact that you're changing that parameter is you would want to set up your own HTTP request directly to Gemini. And so of course you would go to the Gemini documentation and you would basically figure out how you can set up that request. And the way that I got there was I took a look at this documentation and I see if we have high thinking, this is our curl request. But if we go to low thinking, it changes that curl request and it adds this little section down here where we can now customize the thinking level. So that's exactly what I did in this HTTP request right here. Same thing, we have our URL that it gives us, we put in our API key, and then down here, not only are we sending over a text message for Gemini 3 Pro to look at and answer to, but we have the thinking level down here, which is low. And so you can see I can go ahead and run this, and it's going to do the exact same thing as any of these other nodes are doing, whether we go here or whether we go here or here, except for now we know for a fact that we had our thinking level turned to low. But anyways, now that we've looked at the different ways that you can go ahead and connect to Gemini 3 Pro and NN, Let's take a look at the different examples and evaluations that we have set up today. All right, so we're gonna go ahead and start with image analysis. I'm gonna upload an image and we're going to have OpenAI analyze it as well as Google analyze it, and we will compare the results. So I'll kick off this workflow. I'm gonna drag in this criminal justice process flowchart. And what I wanted to show you guys real quick is the prompt for both of these nodes. I'm saying in both of them, describe the process in the image, and that's it. All right, so OpenAI said this image illustrates the criminal justice process from the point of an incident to sentencing, outlining the decision-making steps and possible outcomes. So it says you start with incident or alleged crime, and this is where law enforcement investigates and determines if there is probable cause. It says number 2 is felony referrals or charging and filing. Then we've got summons, bond hearing, we've got case processing and resolutions. And from there, there are several possible paths. Okay, so it understands the structure, but not super detailed. Now if we take a look at Google's answer, we have a little bit more detail here and it's the same type of flow. We start with incident and initial charges. The process begins here, then the path splits into misdemeanors or felonies. The next would be first appearance. If the case is filed, either misdemeanor or felony, it moves into the summons bond hearing, and then it actually goes into detail down here about all the different paths from there. So I know that was just a really quick use case, but hopefully you can see from here how good Gemini is at actually understanding the structure of these flow charts. But anyways, depending on your use case, it might be different the way that you need an image analysis or video analysis in your workflow. So let's say maybe you want to automatically scan in certain maintenance tickets or types of requests like that, or maybe you rent out cars, whatever it is. So let me just go ahead and do a few more examples. This first one is me sending in a picture of some wall damage. So maybe you are, um, a landlord or something like that, and we're gonna have OpenAI and Google once again tell us what they're seeing in the types of damage. Okay, so here's the image I just submitted, and here's OpenAI's answer, which is that the image shows visible water damage along the lower section of the interior wall near the floor. The wall has a large irregular brown stain indicating water intrusion. There's also peeling and bubbling, so the paint has bubbled and peeled away from the wall in numerous spots, and there's potentially mild mold or mildew. So that's what we got from OpenAI. Now here is Google's answer. It says similar things. There's water or moisture damage. There's severe peeling and flaking paint. There's water staining and discoloration. So the overall diagnosis is that because of all this damage that we can see here, there's probably a leak from behind the wall or previous flooding where water wicked up from the floor. All right, so just as another example, let's do a car scratch, and it's gonna be very minor and maybe even a little tough to see. So we'll see if maybe there's a difference here between OpenAI and Google. All right, we got the image on the left and you can see a little bit of scratching down here, and here is OpenAI's response. So we see scratches and paint damage. Noticeable surface scratches are present on the lower part of the rear passenger door, just above the side skirt. A minor dent along the lower rear edge of the rear passenger door, which I guess it looks like there is a little bit of a dent, but not super visible. Now if we go look at Google's response, we have damage located the rear passenger side of the vehicle. There's wheel arc or dogleg area, which is the most severe damage. We've got paint damage and rust, which that definitely looks to be true. And the summary is that the vehicle appears to have sideswiped an abrasive object like a low wall or a pillar, and this resulted in long scuffs on the door and a deeper impact near the wheel well that has since began to rust. Okay, so the point here is that honestly, both of these models are doing a good job at analyzing these images, but with the actual benchmark showing that Gemini 3 Pro is significantly better than all the other models at it, then I would definitely be using Gemini 3 Pro for your automations where you need some image analysis. So we know that Gemini 3 has a context window of a million tokens that it can take in. and its knowledge cutoff is January of 2025. So a lot of times you're gonna have to give it access to like a vector database or just put knowledge in its context window in order for it to actually give you accurate results. So what we're gonna do here is take this 121-page PDF, which is a 10-K of Apple, and I'm just going to copy all of this text and we're gonna go ahead and put it in the system prompt of this AI agent right here. And so now All 121 pages of the PDF live in the system prompt, which the AI agent will look at every time it needs to answer a question. And so what we're going to do here is run an NNN evaluation where we're going to pass through these 10 questions with 10 correct answers, and then we're going to have GPT evaluate how close the answers are to the correct answers. So in our chat model, we have Gemini 3 Pro Preview, and if I switch over to evaluations, you can see that this is being ran right now. So I'll check back in with you guys after all 10 of the questions have been answered by Gemini 3 Pro, and we will see what the correctness score is. And while we're waiting for this, if you don't fully understand the edit and evaluations feature and how that works, I made a full video. I will tag that right up here if you want to give that a watch. All right, so this run just finished up, and you can see that we have a correctness score of 4.6, and that's out of 5, so that's pretty high. The other thing to notice is that for total tokens on average of all 10 runs, it was under 100,000. It was coming in at around 98,000 tokens. And so that just goes to show with a model like this that has a context window of a million tokens, a 121-page PDF like this one is not even a tenth of the way to fulfilling that entire context window limit. So we got 4.6 here as the correctness. I'm going to go ahead and switch this to a different chat model and we're going to see how correct it gets it from there. All right, so Gemini 2.5 Flash came back with a 4.5 correctness, so still pretty good, but not as good. But what we do see is that it was much cheaper and much quicker. So in your scenario, if speed and cost are a huge factor, then you would probably want to experiment more with a Gemini 2.5 Flash for this use case. But keep in mind, this is like a very, very simple example just so we can actually run a few evaluations. So I went ahead and ran one more evaluation with GPT-5 Mini, and you can see it actually did score a 4.6 as well. But really the point I'm trying to make here is that 10 of these examples is really not enough to know which of these models is the best. And it's not as simple as like which model's best, it's which model is best for this specific use case. And that's how you could get in here with your evaluations to understand that. All right, now moving on to this final section down here, we're gonna have Google Gemini 3 Pro actually build us an NNN workflow. And I'm not gonna explain exactly how this flow is actually working. If you wanna see the full video, I will attach that right up there if you wanna check it out. But let me just go ahead and ask this. Workflow to build something for us. All right, so I just shot that off. I basically asked it to create an automation that will take my Fireflies call recording transcripts and analyze it. It will look up the person that I had a call with and do research about them and their business and give me an internal brief about what I can actually do to help them, sort of like an informal AI audit discovery roadmap. So this Gemini 3 Pro NNN builder right now is going to create the JSON. And the reason why I wanted to try this is because we saw some really cool benchmarks with Gemini 3 Pro. About coding. And so it'll be thinking about how to actually structure this workflow, but also building it for us. And then it will create that end-to-end flow and we'll go ahead and look at it and see how good it is. So I'm just gonna open up this link real quick that it provides us, and this should load up hopefully a workflow that is pretty robust. Okay, so the only thing that it really messed up here on first glance was the chat model, and that's simply because this is kind of an outdated prompt that I was using. So let's just say it threw in an OpenRouter model and we're connected. What we've got here is the webhook we would configure to Fireflies to immediately send the transcript to us. We would then feed that transcript information into this AI agent, and this one is prompt to analyze the transcript, use the research tool, find specific pain points, synthesize the transcript, and generate an internal brief. And the brief must have a prospect profile, meeting summary, pain points, and AI audit discovery. We can also see for the research tool, this is supposed to be using Tavily, and it's already filled out this whole request for us. So we've got Tavily there. We also have the actual body set up besides our Tably API key. And it's actually interesting because this is using an outdated version of the HTTP request. As you can see, it's using 1.1, which has been deprecated. But once again, the reason why it's doing that is because I fed it in some knowledge about NNN and it's using some older versions of models. But after it makes that brief, it would go ahead and shoot that off to us in an email. As you can see, I'm just gonna try one more before we wrap up here. Okay, so I just shot off another one that says, create me a daily newsletter that will search sources like Google and Perplexity every day and try to find new discounts on AI tools. If it finds good ones, it should shoot me an email. Otherwise, it has to do nothing, but all of the results every day should be logged in a Google Sheet. So I'll check back in with you guys when this is done and we'll take a look at the workflow. All right, that just finished up. Let's go ahead and click into this link and see what we got for this workflow. Okay, pretty cool. So we've got our daily schedule, we've got our AI Deal Scout, and it also tells us what configuration is still required and the logic flow. Now keep in mind that the actual agent does not have live data. It has static data about NNN Docs, which is why it was probably unable to pull Perplexity and SERP API or something like that. But it does know Tavli, so it pulled that. And it tells us that we still have to configure a few things here. Now the agent itself has a user message saying, hey, here's today's date, search the internet, find results. And if you did find deals, set discounts found to true. And here's what you do if you don't find any. And then in the system message, it says that you are a deal hunter bot. Your goal is to save the user money on AI software. Be skeptical and only report active verified offers. But from there, you can see it also has a structured output parser, and it's outputting discounts found as true or false, and it's also going to output the log summary. And if there are deals found, it will also output an email subject and an email body. Then it's going to link all of that to a Google Sheet, which we would have to set up here, but it will submit the date, the status, and the summary. And then based on that status, it will check over here if deals were found, and if they were, it would actually send us that email notification. All right guys, so there is one more piece that I felt obligated to tell you guys about, and that is when you want to use an agent with Gemini 3 Pro and you want that agent to be able to call tools. It doesn't really quite work yet, and it does, but it doesn't. And let me just show you what I mean by that. So if we're looking at the documentation for the Gemini 3 API, we can see something here called thought signatures. It says Gemini uses thought signatures to maintain reasoning context across API calls. These signatures are encrypted representations of the model's internal thought process to ensure that the model maintains its reasoning capabilities you must have these signatures be sent back to the model with tool calling in order for it to actually work. So basically the flow here is, let's say I ask this agent to send an email. It's going to go to the chat model, understand what to do. It's going to go back to the agent. It's going to go to the tool and send that email. And then all of that's going to come back to the chat model so it can respond to us and say, I sent the email to X and here's what I said. And so that's where we get the error is on that second time we go back to the model. And I'll just show you guys what I mean by that. What I'm gonna do is send off this message that says, send an email to nate@example.com asking if he wants to get lunch this weekend. And what you just saw happen is exactly what I said. The brain calls the tool, the tool works correctly. So if we go to the email, we'll actually see that this did get sent. But then what happens here is we get an error. And when I open this up, it says bad request, error fetching from this URL because function call is missing a thought signature in the function call parts. This is required for tools to work correctly, and missing thought signature may lead to degraded model performance. Now the issue is, when we make this request to Gemini through this native integration, we don't actually have the ability to add in that thought signature or change any of those parameters. And the way that this documentation shows how we should be doing this is, you know, if you went through Python or some other custom way that you can send off these requests with tool calling, you would be able to add that stuff in there. And so first I thought, okay, maybe it's just a limitation of this Gemini node. Maybe we could do that through OpenRouter. I have my 3 Pro preview. I don't see anything in here about, um, thought signatures, but let's just try it again. So I'm going to shoot off that same message. We hit the model, we're gonna hit the tool, it's gonna successfully send, and then we're gonna get another error over here once again. And this time it just says that the provider returned an error. So it's the exact same error going on. And real quick, just wanted to prove to you guys that all of these are being sent. So right here we're getting obviously a message because it's a fake email, but the agent actually is sending the email about wanting to grab lunch. So we know that the tool is working. It's just about the context of what just happened in this tool being fed back into the model so it can actually respond. So like I said, I was doing research on this. I was reading through GitHub forums and NNN forums, and it seems like right now this is the simplest way to put it. It's not working in NNN because the tool calling feature of Gemini 3 requires a special thought signature field in each function call. But N8N's built-in nodes do not add or keep this field when talking to Gemini. For it to work, N8N needs to update its Gemini nodes or workflows so that we can actually include and pass the thought signature each time that a tool is called. And then also just real quick, what is a thought signature? It's an encrypted code from Gemini that saves what the AI was thinking when it made a function call. Then that data gets sent back so it can once again communicate what it actually just did. And so I just wanted to save you guys all the headache of if you've gone through all this other stuff and you've played around with Gemini and you're loving it, and then you decide, okay, I'm going to go ahead and plug plug in Gemini 3 Pro to all of my other agents that I have that are calling tools. I don't come from like a super technical background, so if I completely just messed up and misinformed everyone, then please in the comments, if you know what's going on here and you know how to fix it, let me know. But to the best of my ability, that's what I know. So anyways, I don't want this video to go too long, but if you guys want to play around with this exact workflow so you can just test out some things, then I will leave it in my FreeSchool community, which you guys can go access and download for completely free. That'll be linked in the description. And if you want to take your learning a little bit farther and be supported by over 2,500 members who are building with N8N and building businesses with N8N, then check out the Plus community. That will also be linked in the description. We've got full 4 courses in here. We have Agent Zero as the foundations for beginners for AI automation, 10 Hours to 10 Seconds, where you learn how to identify, design, and build time-saving automations. And then for our premium members, we have One Person AI Agency, where you learn how to lay the foundation for a scalable AI automation business. and then we have Subs to Sales, which is kind of my framework for how I was able to grow a YouTube channel in the AI automation niche and use that to power my businesses. We've also got one live call per week in here, so I'd love to see you guys in those calls in the communities, but that's gonna do it for today. So I hope you guys enjoyed the video. If you did, or if you learned something new, please give it a like. It definitely helps me out a ton. And as always, I appreciate you guys making it to the end of the video. I'll see you on the next one. Thanks so much, everyone.