Gemini CLI 2.0: Can Free Beat $200 Claude Code?

Category: Tech Review

Tags: AICloudCodingComparisonGemini

Entities: Claude CodeCloud CodeCursorGemini 2.5 ProGemini CLIGitHubGoogleMonarch MoneyPlayright MCPVS Code

Transcript

00:00

Google just dropped Gemini CLI 2.0 this week and I know what some of you are thinking. Nope, I only use Cloud Code.

But that's exactly why you should stick around because today we're going to find out just how close Google is getting in the AI agent coding race. So, is Gemini

00:15

CLI finally a contender with Cloud Code? Let's get into it right now.

So, what is Gemini CLI? You may recall that about 2 months ago, Google released this.

And while it was super powerful, the general sentiment in the community was that it was not quite as good as Claude Code,

00:30

but Google has quietly been rolling out new updates to change that. And over the past couple weeks, some huge updates have come in.

We now have native integration between Gemini CLI and your favorite idees like VS Code and Cursor. We have GitHub integrations and GitHub actions just like you have in Cloud

00:47

Code. And we also have better MCP integrations and a new planning mode that handles complex coding tasks a whole lot better.

But you may still be wondering why even use Gemini CLI in the first place. What makes this good?

What makes this different? Well, there are a couple of things that make this

01:03

different from Claude Code and that is Gemini CLI is completely free and open- source. So you get 60 requests per minute and up to a,000 requests per day simply with a Google account.

So, if you have a Gmail account, if you have access to Gemini, you just enter your credentials and you automatically have

01:19

access to Gemini CLI. It also has access to the Gemini 2.5 Pro model, which is a super powerful model with a huge context window.

This is a million context window. This is five times as big as the current window, 200k context window in Claude Code.

And there are rumors that

01:35

Gemini is about to bump this up to 2 million, which is an order of magnitude greater than the closest competitor. It also has built-in Google search which is super awesome because web search is a key component that you use as part of these CLIs.

And I think that Claude Code's web search sometimes doesn't do a very good job and Google search is

01:52

honestly sometimes way better. So, all of that's to say that if you're paying between 20 to $200 a month for cloud code, you should definitely check out Gemini for these reasons because it's free, open source, has a massive context window, a powerful model under the hood, and this is what sold it for me and made

02:07

me want to even create this video in the first place to show you guys how this stacks up to Claude Code. So, how do you install Gemini CLI?

It's very simple. You're going to open up a terminal window and paste in this command, which I'll put down in the show notes.

I've already done that. And once you do that,

02:24

you can open another terminal window and type in Gemini and it will start up the Gemini CLI. And one of the first things you want to do is authenticate with your Google account.

So you can see that there are three different ways to do that. I would

02:40

recommend the first option which is simply just to log in with your Google or Gmail credentials. You can also use an API key or if you have a Vertex AI uh username and password, you can use that as well.

But uh I've already done this. You can just enter O as I as I showed

02:57

you and press enter on option one. It will open up a pop-up window where you can authorize your Google account to connect to Gemini CLI.

So once you set up Gemini CLI, the next thing you're going to want to do is to enable the IDE integration with either Cursor or VS

03:13

Code, whichever IDE you are working in or whichever one you prefer. So right now I'm just going to check the status.

You you can see it's connected to Cursor. The first time you set this up, um you type in forward slash IDE and

03:28

then select the install command. And what that will do is install the integration between Gemini CLI and your IDE.

Now, sometimes the install command doesn't work. It actually didn't work for me.

And one way to get around that is to go to either the VS Code marketplace or the cursor extensions

03:45

marketplace, find the Gemini CLI companion, and install it. And that will automatically connect and integrate your IDE with with Gemini CLI.

Couple other cool features here. Just like with cloud code, you can do forward/mcp to list all the MCP servers you have.

04:01

You can see in this case I have one MCP server which is playright. That will become important in just a second when we do our head-to-head demo, but you can add MCP servers kind of in the same way that you would with cloud code.

Another interesting feature that is a lot like

04:16

cloud code is forward/init. Um, this will generate a gemini.md file which is a lot like claw.md.

This is a way to maintain context in memory in your project so that the CLI agent knows what it's doing, knows what it's done

04:31

before. Right now, we don't have any code in this codebase.

We're going to be generating that code in a demo in just a second and it will generate a Gemini.md file. That's just something to keep in mind.

It's one of the core features that you'll be using if you use Gemini CLI. Another really cool feature is checkpointing.

If you've ever used

04:48

cursor, you know that there are checkpoints that it automatically snapshots or captures that you can roll back to if you make a mistake. If you're using cloud code, if you want to use checkpoints, you kind of have to rely on GitHub or some other system to manage previous versions of your project.

But

05:04

Gemini has this cool feature called checkpointing, which you have to enable. It's disabled by default, but if you start up Gemini using this checkpointing flag, it will automatically capture checkpoints every time new files are written or changed.

So, if you make a mistake, you can roll back. One final

05:20

feature I'd like to highlight before we get into the head-to-head battle with Cloud Code is custom slash commands. This is again a very similar feature to Claude Code, which also has custom slash commands.

You can do the same thing in Gemini CLI. And one of the custom slash commands that I really like is plan.

I

05:36

mentioned earlier the enhanced planning mode. This actually doesn't ship natively with Gemini CLI, but um on the Gemini CLI blog from Google, they recommend this plan command that you can drop right in to your project and have an enhanced planning mode ready to go in

05:53

your code. So to create a custom command, this is very similar to cloud code and super simple.

You can see that in our project here, we have a Gemini folder. We can create a subfolder in here called commands.

And within that we can create a file called plan.totml. And

06:11

this will define the plan/command. And if we go back to the blog here, it tells us exactly the content that we need to drop into that to create the command.

So we will press save and we will relaunch our terminal with the

06:27

custom slash command. And you can see that plan now shows up and we can use it to enhance our project.

All right, now for the part you all have been waiting for. Now that we've got Gemini CLI installed, let's put it head-to-head to the test against Cloud

06:44

Code. Now, this assumes you've already got Cloud Code installed or at least have some familiarity with it.

If you don't, or you don't know much about Cloud Code, I've got you covered with a video right here that shows you everything you need to know to get started. But what I'm going to do right now is give Claude Code and Gemini CLI

06:59

the same assignment and that is to build a personal finance app in the style of Monarch Money. If you're not familiar with Monarch Money, it's a $850 million company.

They have a super popular personal finance app. And I want to see how Cloud Code and Gemini CLI do at creating a simple clone of that app.

I'm

07:15

going to give each of them the same prompt, the same screenshot showing the UI we want to clone, and the same tool, which is the Playright MCP. That's basically going to allow the AI agent to visualize the UI and test it to make sure it works.

So with the playing field level and the rules set, let's go. Okay,

07:32

so I have this detailed prompt here. This basically outlines all the core functionalities, all the features we want to build in this finance app.

And I'm going to come down here in the command line and tell it to execute the instructions that are in this prompt. MD file.

And just like with cloud code, you

07:48

can tag files with at. And to reference the screenshot I have here, you can see this just shows kind of the UI.

This is what Monarch Money looks like. This is what we kind of want the general look and feel of this app to have.

So let's get Gemini CLI started and see how it

08:05

does with the coding task. So, one little interesting quirk of Gemini CLI is that when it sets up all the initial project directories and and folders, even if you've created a folder that you want it to work in, it will just basically create a subfolder sometimes

08:21

with the exact same name as your project in here. And you can see I'm in a folder called Gemini Demo and it created a folder called Gemini Demo.

Kind of annoying, but you know, one of the quirks of Gemini CLI. Cloud code doesn't really do this, but u we'll see how the end result looks.

Okay, checking in.

08:36

It's been about five minutes of work from the Gemini agent. And you can see that it is prompting me to do a few things, but I've turned on down here autoacc, which if you've been using cloud code, you may be familiar with shift tab for autoaccepting edits.

You

08:52

can do the same thing in Gemini CLI so that you don't have to approve every command. And just like with cloud code, it gives you the option of either allowing once with option one or always allowing with option two.

If you click option two, it's going to modify your permissions file within the Gemini

09:09

folder of your project. So, it will know what permissions are allowed for different types of commands.

So, if you want to do something like a YOLO mode, there are ways to do that. Or if you want to approve everything, you can do that as well.

Okay, so quick check-in again. It's been about 20 minutes.

It's

09:26

made some really good progress. And you can see it's only used about 3% of the 1 million token contacts window.

Uh, but we keep getting this error where when it runs the development server, there is a u there's an error with the types that it's created and there's no export for

09:42

account. I've asked it to sort of troubleshoot this, but it seems like it's kind of getting stuck in a loop.

But let's see if it can resolve this and kind of wrap everything up and we can see what it comes up with. Okay, after about 40 minutes, this is what Gemini CLI came up with, which honestly isn't

09:57

too bad for a oneshot. You can see it's got some mock data in here and it has some very basic looking minimal user interface, but you can see how this might be able to be refined if you had more time and more prompts to work with.

A couple of issues I noted when I was

10:12

working with Gemini CLI. Number one, it's a little bit slow compared to Claude Code, although we'll see in a second if Claude Code is much faster.

But generally because Gemini is free and you're limited to 60 requests per minute, I've noticed that it kind of waits sometimes, probably because it's

10:28

not letting you use too many requests at any given time, so you don't break through those rate limits. The other issue I noticed was that as we got closer to the 40minute mark, it switched from Gemini 2.5 Pro to Gemini 2.5 Flash because it used up all the requests I

10:44

had per day. That surprised me a little bit, but I also didn't notice much of a degradation in performance when it moved to the flash model, even though I was expecting it to kind of fail.

The other thing I noticed was that even though Gemini was really good at calling the Playright MCP, it never failed to do that. And it even called it five times

11:01

and went through without much of my input to check the screenshot from Playright against the screenshot of Monarch Money that we gave it at the beginning to keep refining and getting it closer. However, you can see even after five iterations of doing that, it really didn't get that close to it.

And

11:16

we have sort of a not the best looking user interface here, but again, not bad for a oneshot. And I'll just kind of scroll through this here.

And we have some transaction data. Again, very basic, but a lot of the core functionality is here.

I think Gemini

11:33

just struggles a little bit uh with design work. And you can kind of see that manifested by the fact that when I asked Gemini after doing these five attempts to match the screenshot here, you know, why isn't this closer?

It it's

11:49

basically kind of blaming me for having a browser with the wrong dimensions or something. None of which is right.

But I think this is where Gemini falls a little bit short because that front-end work is sometimes really important. I want it to look closer to this mock and it just wasn't quite able to get us there.

But let's see if Claude Code can

12:06

do it better right now. Okay, now it's time for Claude to enter the ring and see if it can do any better.

I've started up Claude Code here and I've entered the exact same prompt in the text box that we gave to Gemini. And I have the exact same prompt MD and

12:22

screenshot that we're using to make this app. Let's see what it can do right now.

And we'll also be checking in to see if if Cloud Code can do this any faster. I suspect it can simply because Gemini is free and and Claude is paid, but we will see.

Okay, so right off the bat, one of

12:37

the key differences is that Claude has a very detailed to-do list. Uh Gemini doesn't do this.

Now, again, you can see that it's very agentic and it plans out a lot of the things it does. And I found it actually with its error handling, it can be very agentic with trying

12:53

different approaches to address an error. But the user is not seeing a lot of these to-dos.

And I like that Claude gives this right off the bat. so you can see what it's doing in real time.

All right, let's let Claude run a little while and we'll check back in in a moment to see how it's doing. Okay, so

13:08

checking in. It's been about 15 minutes now.

And interestingly, Claude Code is running into the same exact error, the same exact type error that Gemini CLI ran into. So this is not unique to either of these agents.

They're all hitting the same exact error. So super

13:23

interesting. Let's see if it can fix it quicker than than Gemini, though.

Okay, so after five minutes, Claude was able to get through the error, which is quicker than Gemini was. And this is what it came up with for the dashboard.

Um, and this doesn't actually look that bad, but I have to say it's not much

13:40

better than Gemini. But just to be fair, because Gemini had about five iterations on the user interface, I'm going to give Claude one more try against the screenshot to get it right.

And we'll see what it comes back with and compare that to Gemini to see who's the winner. Okay, we gave Cloud Code a few more

13:56

tries. And to be honest, the redesign feels actually worse than what it came up with uh to begin with.

So, if you can look here, I mean, this doesn't look too bad, but the dashboard is super cramped on the left side. And it even noticed that and it just kept going around in circles, kind of like what Gemini was doing with Playright, trying to get the

14:14

screenshot of the browser to look exactly like the screenshot we gave it. But neither one was able to actually do it.

So, who is the winner here? This one was a lot closer than I thought it was going to be.

If I had to pick a winner, I think Claude Code narrowly wins this

14:30

one. And I really mean narrow, and only because it was faster.

Look, Gemini is free, but Claude Code took about half the time to create a decent looking app. Now, none of these apps actually look great.

Neither one of Gemini or Claude was actually able to make a clone of the app that I asked it to. Claude Code,

14:47

while generally pretty good at user interfaces, didn't actually make something that looked that good here. Look, its first try was definitely better than Gemini CLI's first try, but as it kept iterating, it kept getting worse.

I honestly thought that this would be a much more decisive victory

15:02

for Claude Code. But while Claude Code narrowly won, Gemini CLI is still very compelling because it's free and it has a massive 1 million token context window.

Google is definitely making things very interesting in the AI coding space. I plan to keep using these tools

15:17

as they evolve. Gemini is shipping updates really fast and there's a rumor that Gemini 3.0 might be around the corner very soon.

But I want to know your thoughts on what you think is better between Gemini CLI and Claude Code. Drop a comment down below to let me know your thoughts.

Thanks for

15:33

watching and I'll see you next