Jack Roberts breaks down the triple Google drop ? Flash 3.5, Antigravity 2.0, and the CLI that replaces Gemini CLI ? and shows you exactly where each fits in a Claude-first workflow.
Posted
3 months ago
Duration
Format
Tutorial
educational
Views
71.1K
1.4K likes
57 · 43
Big Idea
The argument in one line.
Google's new Gemini 3.5 Flash and Antigravity CLI unlock a multi-model AI workflow by letting you delegate fast, parallel-agent research tasks to Google while keeping Claude as your primary reasoning engine.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You use Claude Code or Claude as your primary AI tool and want to integrate Google's faster, cheaper models as sub-agents for specific tasks.
A solo builder or small team running AI workflows who needs cost reduction on multimodal processing without switching platforms entirely.
You're evaluating whether to adopt Antigravity 2.0 or stick with your current IDE setup, and you want a clear verdict on what actually saves time.
SKIP IF…
You're already deep in Google's ecosystem and building exclusively on Vertex AI or Google Cloud — this assumes Claude Code as your primary IDE.
You need production-grade guidance on Antigravity 2.0's parallel-agent architecture — this is a practical integration breakdown, not an architectural deep dive.
TL;DR
The full version, fast.
Google released three things together: Gemini 3.5 Flash, a model positioned as fast and capable at four times the speed and 40 percent of the cost of Pro; Antigravity 2.0, a parallel-agent app locked to Google models only; and the Antigravity CLI, which replaces the deprecated Gemini CLI. The mechanism worth adopting is a multi-brain workflow where Claude Code stays the primary harness and the Antigravity CLI gets invoked as a sub-agent using the agy command, letting Gemini handle multimodal and design-heavy subtasks while other models cover reasoning and code review. Skip Antigravity 2.0 for daily work because it forbids non-Google models, keep the classic IDE, and wire agy into Claude so you delegate by strength rather than commit to one vendor.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Google logo cold open, host intro. Frames the three drops as a stack-level change for anyone in the agentic OS space.
01:05 – 02:33
02 · Gemini 3.5 Flash ? what changed
4x faster than old Flash, 40% cheaper than Pro. Performance-per-dollar positioning. Benchmark chart shown on screen.
02:33 – 06:13
03 · Antigravity splits into two apps
Antigravity 2.0 (parallel agents, Google-only, no terminal, no extensions) vs Antigravity IDE (familiar editor, multi-model). Custom graphic explains the split.
06:13 – 07:44
04 · Jack's verdict ? IDE wins
Antigravity 2.0 is antithetical to best-model-for-best-task. He's staying on the IDE. 2.0 only useful for Google-native parallel tasks.
07:44 – 09:31
05 · Antigravity CLI ? replacing Gemini CLI
Google deprecates Gemini CLI, replaces with Antigravity CLI (agy). Live install demo. Works anywhere: Claude Code, Codex, etc. Shows spend tracking across environments.
09:31 – 11:05
06 · Multi-brain strategy in Claude Code
Using agy inside Claude Code ? Claude orchestrates, Gemini handles multimodal sub-tasks. Live demo: Vancouver temperature query. Note: Claude doesn't natively know the agy command.
11:05 – 13:17
07 · Full workflow demo + Zapier MCP
Claude delegates HTML email design to Gemini 3.5 Flash, uses Zapier MCP to draft the Gmail. Finished Vancouver itinerary HTML shown. CTA to agentic OS video.
Atomic Insights
Lines worth screenshotting.
Gemini 3.5 Flash is 4x faster than its predecessor and 40% cheaper than Gemini Pro — it matches Pro on capability while solving the speed problem.
Antigravity 2.0 is a locked Google-only environment — you cannot use Claude or other third-party models inside it, which is the key tradeoff to know before adopting it.
Google split Antigravity into two separate apps: a parallel-agent platform (2.0) and the classic IDE, and most daily developers should stay in the IDE.
The Antigravity CLI replaces the deprecated Gemini CLI and can be wired into Claude Code as a fast multimodal sub-agent for tasks that benefit from Gemini's strengths.
Using Gemini 3.5 Flash as a sub-agent inside Claude Code gives you Google's speed and multimodal processing without abandoning your Claude-first workflow.
Benchmarks are vanity metrics — the only stat that matters for a coding model is whether it moves your project forward faster than what you were using yesterday.
Takeaway
Get in before the hype window closes.
Format to steal
Every Google drop is a content opportunity ? the window between announcement and everyone's done their take is about 48 hours.
Structure it as three parts: what dropped, what it actually means, live demo with a visible output.
Add your opinion at the 60-70% mark ? not at the start. Let the facts land first, then take the side.
End with a practical thing viewers can do today (install agy, run one command) so the video has a tangible takeaway.
Pre-produce one custom graphic that reframes the story ? Jack's parchment 'Claude Code is the boss / Flash is the worker' diagram is the image people screenshot and share.
Chain to a related video with a 'so next?' setup ? no subscribe ask needed if the content chain is tight.
Glossary
Terms worth knowing.
Gemini 3.5 Flash
A Google AI model release offering faster inference speed and lower cost compared to Gemini Pro, positioned for high-frequency tasks where speed and economics matter more than peak reasoning.
Antigravity (Google)
Google's AI coding and agent environment, similar to Claude Code, that enables multi-agent workflows and IDE integration using Google's Gemini model family.
Antigravity CLI
A command-line interface from Google that replaced the deprecated Gemini CLI, providing terminal-based access to Google's AI models and agent capabilities.
Multimodal
Capable of processing and generating multiple types of input and output — such as text, images, audio, and video — within the same AI model.
Sub-agent
A secondary AI process called upon by a primary AI agent to handle a specific subtask, such as image analysis or web search, returning its results to the main agent.
IDE (Integrated Development Environment)
A software application that bundles a code editor, debugger, and build tools into one interface, such as VS Code, where AI assistants can be embedded to help write and review code.
Resources
Things they pointed at.
09:30toolAntigravity CLI (agy)
12:00toolZapier MCP
13:17channelAgentic Operating System video
Quotables
Lines you could clip.
07:05
“Anti-gravity 2.0 is antithetical to what I think the best way to use AI is.”
Hot take with a big word ? punchy contrast to the hype. Will drive disagreement comments.→ TikTok hook↗ Tweet quote
13:13
“If it's not connected with everything else, we're leaving too much opportunity on the table.”
Clean closer that works standalone ? sounds like a principle.→ IG reel cold open↗ Tweet quote
01:44
“It is fast and capable. Pro was capable but slow. The old Flash was fast but not capable.”
Three-line contrast is inherently clip-ready ? easy to caption.→ TikTok hook↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphoranalogy
Google just dropped anti gravity two point o and the brand new Gemini 3.5 flash. And in this video, I'm gonna show you exactly what's changed with anti gravity two point o and how to use it and the new capabilities that Gemini 3.5 unlocks.
And if you knew, I'm Jack. I built in some molestech startup with the gazillion customers. Now I'm building my own AI startups, and I just shared the stuff that actually works.
So if you haven't already, grab that beautiful coffee, and let's dive straight in. Now Google had been shipping like parrots. And if you're using Energetic operating system, if you're using ClaudeCode or anti gravity, the way that you're gonna use Google's models are going to change.
You can track the usage in systems like this, but you really need to understand what's fundamentally changed. Okay. So let me stop by explaining what's happened here.
Google has dropped three agents, three different updates if you like. We've got Gemini 3.5 flash, which is very, very cool.
It's dropped anti gravity two point o, and it's also dropped something called the anti gravity CLI.
Let me explain first of all about this model, the Gemini 3.5 Flash, and then I'm gonna break down what this means for anti gravity and how we can use this new model in things like Claude Code to do very interesting things with our workflows. It's quite interesting.
A lot's happened. I'm gonna break it all down so get the coffee relaxed, and it all make good sense. So here's the first thing you need to understand about Gemini 3.5 Flash.
It beats pro, basically. So look at this as an instance.
Okay? Gemini three flash used to, from a speed point of view, not be quite up to par.
The idea of the the TLDR would be this, is it is fast and capable. Pro was capable but slow, and the old Gemini three flash was fast but not capable.
That's actually the point of this model. It's supposed to be speed and intelligence. It isn't gonna be the cheapest model that God's ever created, but the idea is it's four times faster than before and 40% cheaper than Pro.
So that's the kind of positioning it's gonna be, and then we'll be releasing 3.5 Pro later. We're gonna put it for test with cloud coding. You'll see exactly how this is capable of in the video, and I've got some stats on screen that you can check out if you want to.
Just to give you some numbers and data here, you got the MCP Atlas. The 3.5 flash new is this guy here, the purple. So you can see how the purple is comparing against the different models on these different benchmarks.
But Jack, we don't want a benchmark queen. We want something that's gonna move our models forward. So we can use this in call today, which I'm gonna show in this video.
But the first thing I wanna talk about is anti gravity. Now if you're watching this video and you've updated anti gravity, you're probably thinking what the hell has happened?
I'm on a brand new computer, so I'm gonna show what this looks like. One very simple thing you have to understand. Anti gravity has split into two apps.
And if you're brand new, by the way, this is just where we do our coding effectively. It could be codecs, it could be the Claude app, it could be anti gravity, these different places. We have anti gravity two point o, which is a bet on what we call a sort of like a parallel agent platform.
In other words, it is isolated. You cannot use Claudine there. You cannot use other models.
It is 100 Google model environment, and the idea is parallel agents.
And then we've got the anti gravity IDE, which is the app that we all come to know and love. Let me show you exactly what I mean.
So I'm gonna shift over now to anti gravity. So when you click on update on your brand new anti gravity, you can come down and select your account. You'll see a sign up page like this.
You can go through the actual setup journey, pick your own specific vibe. We'll go for dark on this one.
Click next. Now this is the app that you're going to see completely different interface than normal.
Couple of things that you need to bear in mind. You can create new conversations here. So I can say, hey there, do me some research.
Tell me what is Glido. Okay. That's the speech to text that I'm using, and you can see it working.
Now this app, Anti Gravity two point o, is Google's bet on where the future is going. Where they believe the future is going is called multi parallelized agents or parallel agents.
And you'll see what I mean. Okay? So for example, when this is doing work, on the right hand side, you're gonna be able to see different things popping up.
And you can do this by clicking this toggle auxiliary pane. And you're gonna see these things here. Sub agents, file changed, artifacts.
Check this out. Okay? I'm gonna say, spin up for me three sub agents to do some research on different speech to text ideas and what the best methodology is.
Alright? So you can do this and you're gonna see now these sub agent spawn ops. So the idea is that the panel on the right hand side, okay, we can actually view and see everything that's happening.
It will tell you different files. And I might actually just come back over and say something like, do me a favor, also create for me five mini documents on my home page covering technology, market opportunity, tech stack, and anything you might think is relevant in building one of these myself.
So I'm just giving it some random prompts. Right? Let's go ahead and do this so you can see how it's working.
Now, this is the point of Anti Gravity two point o, that you're supposed to have these sub agents. And you can click on these sub agents and have a look at what they're doing. Right?
That's the whole idea. And the UI here is trying to build around it. Right?
Which is interesting. But you are also losing quite a lot in this interface because what we can't do now so you can't send a message to Sub Agent, for example.
Right? We cannot at all bring in Clerkout.
You can bring in no other model. This is pure Google City.
You are literally locked and loaded talking to Google City. That's basically it. One other cool thing that they've done here though as well, I wanna call out, is you can go to schedule tasks and you can create a schedule task.
This is exactly like we did in the core code app, in the Codex app. You can basically create new tasks and say, hey, I want you to go ahead and do x y z.
You can effectively do all that if you want to. Again, you can come down here, look at your projects. And as you can see, just like everything else, you can align and sign different things off if you want to.
So this is the new app. This is the direction they're heading into. But a lot of people are like, well, dude, where's the terminal?
How the hell do I get my terminal? There is no terminal in Anti Gravity 2. So to actually get the terminal, what you're gonna do instead is head over to this website, which is this one here.
Of course, I'll put the link for you down below. And very cheekily and near the bottom, okay, you're gonna have anti gravity IDE. And literally guys, come down and click on download for Apple or Intel if you're on that one if you want to, and just download that to your computer.
Then once you do that guys, you'll see almost like the dark side and the light side of the force. We have the black and we have the white. And then you're gonna come down here is literally begin.
You're gonna sign in with Google just in the same way. Then we begin this and we can sign in and see everything we're doing. We're happy with review driven development.
Again, you can just decide what you want to, custom configuration about how you want it to go. Ideally, you don't want it to be asking for a million permissions all the time, but again, you can configure this to your own specifics. And once you've done that and you sign in, you will see an old friend.
An old friend that we're used to using, and just like before, we've got our extensions. Because remember, in anti gravity two point o, you don't have extensions. Now, this comes on to my recommendation and then the new unlock with Gemini 3.5.
So what do I think about this? Honestly, the direction travel, this anti grav two point o is antithetical to what I think the best way to use AI is.
And what that is is different models. In other words, best model for the best task at hand. And since you can only use Google models in this particular harness right now, my current view is that I'm not going to be using this new anti gravity harness because I can't use Claude Code.
I can't use any of the other models that I want to. Maybe I wanna use TrapGVT 5.5 or DeepSeek.
I'm unable to physically do that in this harness. And for that reason, I don't think Gemini is at the point by itself where it eliminates need for everything right now.
So I'm gonna be using the classic IDE to build everything out. That's the distinction I think is really important to understand. And so you can use anti grabs two point o just for agents.
So it's called use cases. Maybe you have one task that you're happy for to do with Google and you want multiple agents. You can do it in this particular environment.
I think it's an interesting move splitting them out, but honestly, I think like the IDE crushes because we can just bring in all these beautiful agents to it. So you can see, anti gravity two point o can dim things. It can spin up three to five parallel agents, schedule tasks, voice input, etcetera.
And then the IDE is the familiar thing that we're used to using. And this brings on to the second big change that's gonna affect the way that you work with Gemini models. So we used to have this thing called the Gemini CLI.
We loved it. It was great. It just let us tag in and use Gemini in cold code and codecs.
Google is deprecating it. It sent it on its retirement plan. We're throwing it a big beautiful retirement party.
We even got a cake, and we're sending it away. And they are close sourcing it and turning it into the anti gravity CLI.
And let me show you exactly what that means and how you use it. Beautiful. And I don't have the anti gravity CLI installed on my computer, so we're gonna do this together.
I've opened up a terminal in anti gravity, which we're using because we're talking a lot about anti gravity today. And just gonna enter in that code. And this will install the anti gravity CLI.
And what this effect would mean is we can use this anywhere. We can use it in cold code, we can use it in codecs, anywhere our heart desires. And crucially, this is a really cool thing, when we actually come over and we build out our operating systems, it shows things like our spendage and cast across Claude, codex, and anti gravity, something I built in the community.
I've done very various different videos on this. We can actually get a full overview of how many conversations we had, what's our spending, what's our limits, and get various different insights based on how we're physically building and using everything, which is really important. So I come back over and it now says that it's installed.
How you actually physically use it now is you're gonna enter the code a g y, like so. Okay? When you enter a g y, you'll see anti gravity begins to pop up.
And then, you know, just classically, like everything else, you can just pick the different themes that you want. You can agree, go down, select everything that you want to. And now basically, what this is basically Gemini 3.5 flash high.
And then you can chat with it and say, hey, what's going on? And now we're effectively talking to Gemini. So the key unlock here is that we're using what we call the multi brain strategy.
So we're gonna use Claude. We're gonna use the best ChatGPT 5.5 model and this new Gemini. Now here's the big unlock with this new model.
It's fast and intelligent. Not necessarily super cheap, but it's, like I mentioned earlier, speed and intelligence. But the cool unlock here is that Gemini, using this new anti gravity harness, can spin up multiple sub agents.
And so for example, this Claude agent here can spin up Gemini and then Gemini itself can spin up. You know, it may say, hey, I think I need five different agents to do this and it itself can spin up individual paralyzed agents.
Let me show you exactly what I mean. So first of all, if I come over to ClaudeCode, we can see this working in practice. Hey there, what I'd like to do is use the anti gravity CLI and just check for me what is the average temperature in Vancouver in July 2026 expected to be.
Alright. Something completely random, but we're just looking to see if we can use the anti gravity CLI. And just like that, it's completed it.
By the way, Claude itself isn't aware of the anti gravity CLI, so you need to use expression a g y. And as you can see, it's run the query with anti gravity.
It's asked the model and it's come back. So let's put this for test by giving it an actual challenge. Awesome.
What I'd love me to do is build for me a one week travelers itinerary for Vancouver for somebody coming from Dubai.
And what I'd like to do is create a beautiful, gorgeous HTML. I want you to delegate the creation of this email and different research components to it to the anti gravity model, the Gemini 3.5 Flash, so we can get some speed and some beautiful graphics. And then I'd like you to basically save it as a draft to myself in my inbox.
Once you do the research and then delegate that to go and design all the components and make it look gorgeous. Now typically speaking, we wanna bring in Gemini. When you're thinking multimodal, like different ideas, we can tag in Gemini.
But when we're using things like code review, we bring in GPT and we can tag in the different models based on what we're doing. But this is essentially how we're gonna integrate them. And then for the actual connections in terms of how I actually send these emails, the one I really like to use, obviously, you've got these connectors.
It looks a little bit different in every single app that you have. I really like using Zapier MCP. If you don't have a direct connector for it within, say, Claude for instance, or you can't find them, there's a couple of options.
You can add custom ones, but actually what you'll find I mean, use something like Zapier, you can build up these individual MCPs that have very specific accesses and we can use this to connect to anything we want to. For example, this is the only place where you can go and get a school MCP, if you were doing that for example.
So this is like Willy Wonka's marvelous emporium of different activities and things like that. Then you can literally come down, add this as a custom MCP, and it just enables us to get access to loads of stuff. You don't have to use it.
I just find it super easy to use and very helpful. And just like that, it's now created the Gmail draft tool. It's kicking off the research, and it'll begin building everything for us.
Beautiful. And then we've got this fantastic little design here that's come down, and we can see we even got these little mini icon type of thing that's been added in by Gemini, which I think is great. And we have our full beautiful thing.
I pulled this up in HTML, so I can actually show what this looks like. Pretty cool. I think this is a pretty sweet design to get started with.
Now the anti gravity CLI is one thing, but if it's not connected with everything else, we're leaving too much opportunity on the table. So next, we need to slot it in with our argentic operating system, which we can learn in this video right here.
The Hook
The bait, then the rug-pull.
Three drops landed at once and if you're building with Claude Code, Antigravity, or Codex, the way you use Google's models is about to change. Jack Roberts, who has been tracking the Antigravity ecosystem for months, breaks all three down before the algorithm moves on.
Frameworks
Named ideas worth stealing.
10:00concept
Multi-Brain Strategy
Claude (orchestrator/brain)
Gemini 3.5 Flash (multimodal worker via agy)
GPT-5.5 (code review)
Use best model for best task ? Claude decides and delegates, Gemini executes multimodal and speed-sensitive work, GPT handles code review.
Steal forAny Joe content framing multi-model agentic OS ? this is exactly the MCN+ pitch of 'own your stack' applied to AI models.
10:12model
Claude Code is the Boss. Flash is the Worker.
Claude plans, decides, and delegates. Gemini 3.5 Flash executes parallel sub-tasks at speed. Clear command hierarchy for multi-model workflows.
Steal forKilling Excuses / Mod Boss framing ? one brain running many workers is a monetizable concept.
CTA Breakdown
How they asked for the click.
VERBAL ASK
13:13next-video
“So next, we need to slot it in with our agentic operating system, which we can learn in this video right here.”
Smooth internal link ? no hard sell, natural content chain. No subscribe ask.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
A five-level walkthrough of building one locked brand system in Claude Design, then generating a burger-brand website, a 3D scroll showstopper, and a matching animated video from that same system.
Jack Roberts open-sources the seven-step Claude Code skill that chains Fable 5, Higgsfield, and a GPT-5.6 juror into an assembly line for animated product websites — for about $10 a build.
A five-level blueprint for turning one frontier model subscription into a personal operating system that remembers everything, debates itself, and keeps working while you sleep.