I Tested 4 Models for n8n So You Don't Have To

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: I Tested 4 Models for n8n So You Don’t Have To · Channel: n8n · 07:27 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI evaluation and benchmarks, AI content generation, AI models

Summary

The video compares four models used through the n8n Assistant to build the same SEO and AEO audit workflow, measuring build time, cost, and output quality. The models are MuseSpark 1.3, Claude Opus 5.5, GPT-6 Astra, and DeepSeek v4.1 Flash, all connected through OpenRouter. DeepSeek stood out for its cost and styling, while GPT-6 Astra ranked highest in the creator’s own workflow audit. The creator emphasizes that this was a single test, not a benchmark, and that results may vary with different prompts.

Key points

The same prompt asked each model to build a URL-based SEO and AEO audit tool that returns an overall score from 0 to 100.
The creator used a Google article about AI and GEO as guidance and kept the prompt relatively generic to test how models used available tools.
Anthropic’s model confused the Appify API key with the TypeSafe API key, while OpenAI asked for its API key before testing.
DeepSeek recognized the available OpenRouter API key and used it to access the judgment model.
The creator tested the workflows on articles considered bad, average, and good, and compared their audits and styling.
Opus and OpenAI fetched additional context beyond the provided Google article; weaker models generally stayed with the context from that URL.
DeepSeek cost $0.26 and performed well; Meta cost $0.35 and had weaker styling and auditing in this test.
The creator’s own workflow audit ranked GPT-6 Astra first, followed by Opus 5.5, DeepSeek, and MuseSpark 1.3.
The creator cautions that this was a single test, not a benchmark, and results can differ with other prompts.

Chapters

Start	Chapter	Summary
00:00	The comparison and test prompt	The creator introduces the comparison, explains the cost and capability context, and gives the shared prompt for building an SEO and AEO audit workflow.
01:37	Setting up and testing the models	The creator describes the self-hosted setup and compares how the four models handle credentials and initial workflow tests.
03:11	Comparing audit results	The workflows are tested on articles with different quality levels, and the creator discusses the models’ scores, audit details, and styling.
05:25	Costs, timing, and workflow quality	The creator compares the models’ use of context, costs, and build experience, then ranks their workflows using a separate audit.
06:57	Final ranking and caveat	GPT-6 Astra leads the creator’s workflow audit, with a reminder that the results come from one test and are not a benchmark.

Quotes

by the end of this video, we’ll know how long it took and how much it costs — 00:00

this model has been great so far. — 02:39

it’s the only one that recognized that I had an OpenRouter API key. — 03:11

there’s DeepSeek that cost us $0 .26, and it performed pretty well. — 05:55

this is a single test. This is not meant to be a benchmark. — 06:57

Concepts and entities

n8n, n8n Assistant, OpenRouter, Jev, GPT-6 Astra, Opus 5.5, DeepSeek v4.1 flash, MuseSpark 1.3

Full transcript (original language)

00:00 I just use these four models instead of the n8n assistant to build me the same automation. So by the end of this video, we’ll know how long it took and how much it costs to build with each one of them. And of course, along the way, you’ll see exactly how is the experience building with each. First thing I want to show is this table. I got the data from artificialintelligence .ai, and it helps us understand what to expect from each one of these models. Because despite one model being dumber than the other, the cost per task for that model seems to be much, much cheaper.

00:31 But then again, this isn’t a benchmark for the n8n Assistant, nor will we be actually building that benchmark on this video, but it’ll help you understand what to expect from weaker models and also how it performs with better models. Obviously, my prompt will be the same for all four, which is build an SEO and AO audit tool, It should start with a form where someone pastes a URL. The workflow then goes to that page, audits it for SEO and AO, and gives an overall score from 0 to 100. use jev for the judgment calls in the audit use whatever else you need for the rest i also gave it some additional details on how it should go on

01:04 and fetch for what is a successful seo as well as aeo The only hint I actually gave it was the URL to this Google article that successfully explains what is AO as well as geo, which, I mean, are usually the same thing, but they lead to these kind of hacks that really don’t. aren’t efficient they don’t work and this page explains to the llm that it should not be looking at that but as you can see in the prompt there’s not much details that’s because i want to actually test these models and see how efficient they are in using the available tools to go on and build our

01:37 automation all right let’s get to it first things first everything is running self -hosted so i’ve created a container for each different provider slash model just in case you’ve been living under a rock and you don’t know how to configure the n8n assistant just come over to settings click on n8n assistant and in here you should be able to select your model and as you can see in this window i have muse spark 1 .3 contributor as the model selected in my other window we’re using cloud opus 5 .5 by the way all of them are connected through OpenRouter this is just to make it

02:07 easier for us then we have openai gpt6 astra And finally, DeepSeek v4 .1 flash. So there you go, pasting the prompt into each different window. Neither of them got anything exactly wrong aside from the credentials. So let’s start from Anthropic. So far, this is actually the most disappointing model, especially because of how pricey it is. It ended up confusing my Appify API key with the TypeSafe API key. So all I had to tell it was to remove the Appify key from there and actually

02:39 use my TypeSafe API key. which is kind of embarrassing. Now, moving over to OpenAI, because it was smart enough to ask me for the API key before actually testing anything, after I gave it that API key, it went on, did the live test, everything, apparently worked smoothly. And honestly, this model has been great so far. Moving on to meta, all I had to say was test the workflow. It apparently did. And That’s it. Now, my favorite model till now has been the DeepSeek model, just because it’s the only one that recognized that I had an OpenRouter API key.

03:11 And because I had that OpenRouter API key, it used OpenRouter to use Jeff. So I mean, doing this is definitely a thumbs up. But now let’s do the ultimate test, which is sending over articles that we already know by a fact that they are bad. So this page is considered bad SEO and AO wise. This one is average. And this one is considered a very good article. All right, so these are the results after executing the workflows. By the way, this is inside of the forum. If we wanted to send this through Slack, Telegram, whatever, it could just be it.

03:45 But also because I wanted to show how it would deal with the styling of it like this. As for OpenAI, it gave a 63 out of 100, which isn’t good considering this URL does not have a good SEO and AO. And the rest of its auditing is, I mean, I feel like it’s a bit generic. There’s a lot of indicators that NAI wrote it. The overall styling isn’t that great. And honestly, it did give us a lot of information on how to improve that specific website. So it’s

04:21 better styling, but it failed to give us more context on what to do. From all of the results, I feel like this is the one that played it the safest, but it’s also the one that answers not so much like an AI. I also think this score is quite good. I like that it separated the SEO score from the AEO score. So I’d say it did pretty well. Now let’s take a look at DeepSeek as well as Meta. First, DeepSeek. I loved this styling. I’m almost scared that I’m being too biased towards this model with none reason

04:51 whatsoever, but this score is pretty solid. The overall styling, like I said, is pretty great. And the observations are pretty similar to the ones from Opus 5 .5. Now, last but not least, Meta, it gave it an 8 out of 100. So it seems like the harshest model. And… The overall styling is the worst, as well as the auditing. I feel like the AI part is okay, but SEO wise, it really wasn’t that great. All of the models successfully read the Google article.

05:25 One difference I noticed is the weaker models just stayed with whatever context it got from that URL. And Opus, as well as OpenAI, went on to actually fetch for more context and add more auditing angles. I had each workflow test all of the websites. Just to make sure the video didn’t get boring, I added all the tests that I made to this table. The green scores here are numbers in between. So meta is 8, while OpenAI is 63. So the ones in between are 35 and 17.

05:55 the Anthropic scores are mostly fine, while the Meta scores are either too high or too low. I also checked OpenRouter for the last cost, and this is really surprising because Meta cost us $0 .35, which is okay, but it didn’t perform that well. But then there’s DeepSeek that cost us $0 .26, and it performed pretty well. Like the styling was on point. Even the time that it took to build everything wasn’t that high, and it didn’t have issues like Meta did. As for OpenAI,

06:25 in building it, but it wasn’t the one that performed the best or styled it the best. But then we have the Anthropic model that performed reasonably well, but the cost for that was just too high. Now, just to be clear about this table, I changed it over to the model’s names instead. Now, the last and probably one of the most important scores is about the workflow and actually what it produced, because these scores is just an average from whatever it came up with. But we also want to understand how actually

06:57 wanted to add human in the loop or just include Slack as how it would message me. Those are all iterations that would end up costing you more money and time if the core automation isn’t well built. So I ran my own auditing system on top of those workflows and the best performing one is GPT-6 Astra, followed by Opus 5 .5, then DeepSeek, and lastly, MuseSpark 1 .3. And just to be clear, this is a single test. This is not meant to be a benchmark. Models will certainly behave differently

07:27 especially when your prompt is as generic as the example used in this video. That is it for today. If you want to see more tests, let me know down in the comment section, leave a like, and subscribe for more. See you in the next video. Till then.