I Tried NEW Cursor 3: They (try to) Change The Game

Hello guys. Cursor launched a totally new version, cursor 3 and in this video I will show you how it works.

I would call it a mini revolution because they're changing what Cursor as a product even is. So if you see this view and this is how it works with Dark mode on my local machine and I will show it to you in action.

But if that visual multi agent environment looks familiar to you because you've seen it somewhere. Yes, Cursor now is very similar to Conductor for multiple agents.

So here's the screenshot. And by the way, conductor recently raised 22 million serious A so they're getting serious.

Also there are other similar tools for multi agents like SolarTerm by Aaron Francis, like Polyscope from Beyondco team from Laravel Community.

Also recently I saw Superset. So multi agent environments are pretty probably one of the most popular products these days.

And also don't forget Codec app is also looking similar of course limited to just GPT models, but still also multi agent.

I'm actually surprised that cloud code with Anthropic haven't released something like that yet. Maybe it's in the works.

We'll see. So with this move, cursor 3 becomes not a versus code fork anymore, not IDE, but also multi agent environment for now, not the default view.

Because when you launch Cursor you will see that at the bottom. At least it was for me. Here's a screenshot.

New window for running parallel agents, they call that. One of the AI reviewers, Peter, is saying that that should be the default view.

But for now the Cursor team is more cautious and is waiting for community feedback. So let's see how it works in action.

So now you will see Cursor three features in action and Agents. But in addition to those features, the pricing still remains a very hot topic and that still the differentiator of whether someone wants to use cursor 3 or other multi agent environments.

We'll talk about the pricing at the end of this video and I will show you how much it costs for running Multiple agents and Cursor for now the features.

So as you can see on the left you can open multiple workspaces, multiple folders on this Mac or Cloud or Connect SSH and in each workspace you may add Agent and just start prompting.

And this is my example in one of my local projects this was the result of this simple prompt of changing the text.

This is a typical Cursor result. Then I click the button to commit, which happened on a separate branch Cursor as you can see the branch name.

Then Create pr created the PR on a real repository here this is what was done. So the commit message actually says made with cursor.

But this is different from cloud code which happens to be the co author of the commit. So this is how it works in general.

And now I will try three different different workspaces with three agents but for the same prompt. So I prepared three folders with new Laravel installation Cursor Opus Cursor GPT and Cursor Composer.

Those are identical repositories. And I will launch the same prompt and we will compare the pricing and how it generally works with multiple agents.

So open Workspace one more and Cursor Composer gives us three workspaces here let's remove that workspace for AI coding daily remove from sidebar.

You can do it this way. Let's remove that one as well. So now we have three workspaces and let's launch three identical prompt but with different models.

So this will be Composer 2, their latest model. So I'm launching that one and let it run new agent. Then for GPT, same prompt but I will choose GPT.

I probably will stop call it codecs. It's GPT now without codec suffix. And then in the opus as you can see, those are running already.

And for here I will choose Opus 4.6 and also I want to try myself. How much will it cost to create a few cruds with latest frontier models like Opus and codecs?

So here's a to do and you can switch between the agents between the prompts. So yeah, each of them is running and I will wait for them to finish or at least one of them and we'll see the result.

And we're three minutes in and Composer two as expected, was the fastest because well, it is the fastest model worked for 3 minutes and 21 seconds.

And here's the summary with passing tests. Does it actually work? Let's launch it in browser. So this was the project on top of default Laravel Starter Kit.

We have categories and posts. Probably we need to run NPM run build, which Cursor didn't do in this case.

So I should probably open a terminal. I'm not even sure how to open the terminal in this view. So I'll just open Ghosty and run NPM Run build here.

So this is actually a good question. If I am in that agentic mode, I kind of lose the view on the files and the terminal by default.

Probably there's a button somewhere to open the window. Oh yeah, I see terminal here. So I could open the terminal here.

There's a button on top. Right. But yeah, I already did that in the separate terminal. And yeah, it got a bit better stretching the table.

So does it work? Yep, it seems to be working with Composer 2 now. Other models are five minutes in with GPT still in progress.

And for Opus it's like much slower. One out of seven completed. And this by the way, doesn't mean that Opus is slower because Opus is running in cursor.

So maybe cursor is the reason for slow behavior. This is why personally I'm a bigger fan of launching the model in its native environment.

So cloud code for anthropic models and codecs for GPT. But cursor may be a middle ground for many models.

And some people actually report that in cursor, cloud code models are doing sometimes better. So again, let's wait for them to finish or actually meanwhile we can take a look what we can do with Composer.

So here's the button, Create, Branch and Commit. If I click Review, what does it do? Yeah, so we can take a look at the code that was generated.

Great. And then the button, the same button here and then on the top right, Create, Branch and Commit.

So I've created a GitHub repo for this testing and let's try to click Create, branch, Commit and push and I will show you what happens.

So you're creating a new branch, generating a command, and this is actually you're paying for that by the way.

So if you don't push to git yourself for all those operations, they cost tokens. And the pricing for Cursor specifically may be, as I said, a pretty important factor.

But we can see the commit now and the push and we can create a PR. We can create a draft PR or a regular PR.

Let's do create PR. And will it take us to GitHub automatically? We'll see. Okay, so we do have a pull request and here's the link and let's see what's inside.

So summary database, how to verify. Again, made with cursor. And the branch name is in cursor. And yeah, from here you may merge or do something with that change.

And while I'm talking, one of the models is finished, which is GPT 5.4, worked for 8 minutes and 50 seconds.

And by the way, you can see the review here. Thousand lines. And Composer was, from what I remember, less than that, but I don't see that anymore, that button.

But from what I remember it generated less code, but let's see the review and you can actually compare the code here.

So here's the category controller with index in GPT and here's the same category controller, much more simple or actually it's just horizontally aligned so it's not much worse or that different.

But of course it's a very simple crud so it's not really that much to compare. It's just like with count for example was with GPT model.

So it figured out to show categories with the amount of posts in the table. And then while we're talking opus is finished already.

And review also with count here by the way, which composer didn't do. And this is probably one of the examples of composer being fast.

But sometimes being fast is not really deep enough. But it may be good for regular day to day operations.

Now let's see if it works. So here's cursor GPT test again without NPM run built probably. And for that let's actually try Terminal in action.

This is not in the video but I will scroll it down. So here's the button for terminal and here I will run NPM run build and now I will refresh.

And yeah, this is the table with weird misalignment here. But does it work? A bit more interesting layout with the card but it does work but also without the icons here on the left.

And here's the result by Cursor opus which didn't work because it didn't execute the migrations automatically for data database.

So as you can see at the bottom of GPT it ran migrate successfully. And for opus probably I would guess it would tell us to run migrations.

Nope, there's no instructions to run migrations but I still know I have to do that. So PHP Artisan migrate.

Now we have those tables and it will be probably good. And we also need to run NPM run, build and refresh again.

And we have the categories

here so it's comparable results. Again it's a small crud, but it kind of works. But as usual with every AI agent in every harness it may be different.

Some NPM run may be not executed, some migrations so you still have to verify supervise and stuff like that.

But my goal here was to show Cursor in multi agent and multi workspace environment. So here's how it works in general.

One more thing and important differentiator is cloud agents by cursor. So for example you can create an agent and choose not this Mac but cloud for that agent and it will execute that query that prompt in cloud environment.

Here's the page on Cursor about cloud agents. It will be probably similar or identical if you go manually to cursor.com agents and run the prompt on one of your repositories.

But let's try that from Cursor. For example, let's ask it to change something on the homepage. Change.

Let's get started to hello on homepage in welcome Blade. No, it doesn't find the file for whatever reason, but maybe it will help.

So Opus is still the model, but it's working in the cloud, I assume. Let's see if it actually works. Send doesn't work.

I'm not sure why actually and I don't know how to debug. I haven't run it executed before. So maybe there's a bug in cursor.

But just to demonstrate how it may work, let's try to do that manually here. So I will choose the same repository and same prompt and I will choose Opus 4.6 hi, for example and let's see what happens in the cloud.

So it's setting up cloud environment and we will see the pricing. I'm curious myself how much will it cost?

So let's click and see what's inside. It will show probably roughly the same thing as you would see in Cursor locally is just that setting up that environment will take probably some time and maybe also it is slower because a lot of people are trying that at the moment as cursor three is pretty new.

So yeah, cursor dependencies 36 seconds out of curiosity, what are cursor dependencies? So Ubuntu so it's preparing virtual machine but from my experience with those virtual machines is sometimes it doesn't have the necessary tools for specific tech stacks.

So probably it will not have PHP or something for Laravel already. I'm not sure. We'll see. Usually the main question is will it be able to run automated tests?

But the change is correct successfully in the Blade. And yes, it's starting to write automated test, which is great.

It's in the cloud MD and agents MD the instruction to have the automated test but will it be able to run them in that virtual environment?

Okay, so it's updating that test with the new hello message run homepage test. So this is. I'm very curious about PHP command Not found exactly what I was expecting.

But what it will do now find PHP binary, no output, no php. Right. Oh, and it will install it so it didn't happen before when I tested virtual environments like a few months ago.

But even if it does succeed, like see what is the hassle with the cloud agent and cloud environments, it's probably much faster to do that on your local machine and much cheaper.

And you would probably use cloud agents and cloud environments, probably for the cases where you want to prompt the agent and don't even have the laptop with you.

And the agent would run in the cloud somewhere. But that may be pretty expensive. We will see the price when it's done.

One more thing with the pricing about those cloud agents. So in the documentation of cloud agents, the kind of warning is that models always run in max mode.

What is max mode? In cursor it extends the context window, which is basically meaning consumes usage faster and models API rate.

So cloud agents are even more expensive as I understand than running that locally. And what do we have here now after one minute.

Oh, it's interesting. Check PHP install progress. So it is installing PHP with one sub agent and then another sub agent is waiting and checking sleep 60.

So this is interesting decision. Kind of off topic in this video, but I wanted to show it in action. So there we go.

And there we go. In a few more minutes we are done with php. Now composer dependencies and then running homepage test failed, Missing app key and exception.

Yeah, I know. It needs to run php Artisan key generate and now both tests pass. So it did succeed in the end.

Now let me commit push and create the pr. So this is what Cloud Agent will do for you. It automatically creates the PR and this is done.

It is saying that it worked for five minutes, but I'm not sure. To me it felt much longer. I didn't really calculate.

But we can go to the pull request on GitHub and there we go. So in this case it's different in terms of who is the author.

So cursor is the author in this case. And yeah. Now let's get to the most interesting part, the pricing.

Let's look at my cursor dashboard. And there we go. First, the usage today. And this is probably the main thing that you need to know.

For all those prompts combined, the daily total usage is $5. From my $20 plan, it is said to be more than $20.

It's usually like $40 or something. But this is approximate and this is actually the problem with cursor pricing at the moment it's still the same on that rate.

So $5 for those prompts, which means roughly eight more prompts like this or eight operations like this.

And I'm done with monthly usage. Not daily, monthly. So for the same 20 or $25 per month subscription on Cloud code or codecs I would get much more usage.

It's not

and if we took a look at individual operations here. So here's the cloud agent with opus 4.6 which took 700k tokens for just the change of the text.

And for example local opus took 2.8 million tokens which was okay, a bigger operation for a few cruds and Composer two took fewer tokens.

It doesn't say the price specifically, but you can probably calculate that cloud operation of opus took like $1 or a bit less than $1.

Just changing the welcome text. And Composer is cheaper, but not that much cheaper. So I'm not sure what is Composer too fast is doing that?

Probably it was kind of sub agent, but in total 95 cents for composer for still one prompt for a few cruds in a simple Laravel project still in my opinion too expensive.

So as I said, the price will still be probably the main differentiator of people not using cursor even with multi agentic environment which they seem to be good at.

So for example Richard's tweet here comparing cursor to Conductor and the main differentiator is this 10x.

This can use codecs and cloud subscriptions. Inside of Conductor there's a gray area if you can do that because anthropic is pretty famous for taking down the usage like in open code you cannot use cloud subscription and stuff like that.

So I'm not sure how it works in Conductor. So for example, in another tool, Solaterm, how it's doing that, it's actually launching the terminals of cloud code or open code or Gemini.

So it's not working with subscriptions directly, it's working with terminals. So that is probably legal.

So we'll see what the future brings with can we use cloud and codec subscriptions in other multi agentic environment.

For cursor it seems to be the same old pricing model. So you pay the cursor and then cursor pays almost API pricing to the LLM providers.

But if we take the pricing argument aside, strategically, directionally, I really like where cursor is going because I think this is the future how we will work with agents, with coding agents we will prompt and then review the code and then make some manual changes when needed or even re prompting the agents.

And also at any point you can switch to regular cursor and use that as ide if you want. So it's still doing both, but agentic first environment seems to be the direction.

So yeah, what do you guys think about cursor 3 in general and about the direction of multiagentic environment where seems to be all going for?

Let's discuss all of that in the comments below. And if you want to see more of my AI coding experiments with various tools, go to aicodingdaily.com There's a list of articles and videos.

Some of them are premium for deeper dives. And you can also subscribe to weekly newsletter which I send every Wednesday with news from the community and the latest trends.

And also I show various prompts and tools that I find on Twitter over that week. So subscribe to that newsletter.

That one is absolutely free. That's it for this time and see you guys in other videos.