6:23:04Nick Saraev · Tutorial
Vibe Coding Full Course: Gemini 3.1 + Antigravity
Six hours, five shipped apps, and one uncomfortable conclusion: the building is now the cheap part.
February 27thThe argument in one line.
Coding agents no longer require a credit card or subscription: FreeBuff trades ad views for daily model hours, including a 20-hour-a-day free allotment of GLM 5.3 Flash, narrowing the access gap between people who can and can't afford frontier AI pricing.
Read if. Skip if.
READ IF YOU ARE…
- You want to try AI coding agents like Codex or Cursor but can't justify $120-720 a year in subscription costs.
- You're curious how an ad-supported monetization model works inside a developer tool rather than a media or gaming app.
- You want to see live, unedited output from a free-tier model on a real one-shot coding task before trying it yourself.
- You're in a country or income bracket where a US-priced AI subscription is a genuine barrier to entry.
SKIP IF…
- You need production-grade code output — both demos here are toy one-shot builds (a bio page and a walking stick figure), not a stress test of coding capability.
- You're looking for a rigorous benchmark comparison between FreeBuff's models and paid competitors — this is one person's first impressions, not controlled testing.
The full version, fast.
FreeBuff is a coding-agent app that charges nothing upfront: instead of a subscription, users watch ads and draw from a daily allowance of 'Freebucks' good for hours of use across models like GPT-5.6 Luna, DeepSeek V4.1 Flash, and GLM 5.3 Flash, the last of which alone offers 20 free hours a day. Nick Saraev signs up with no credit card, opens two parallel build threads, and prompts one model to build a personal one-page bio site and another to build a simple stick-figure walking game, both of which render working results within minutes while ads surface quietly at the bottom of the screen. His conclusion: the free models aren't frontier-tier, but they're good enough to code with, and giving people without money access to any AI coding capability is worth taking seriously even though the ad-funded business model's long-term economics are still unproven.
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.
Create a free account →Where the time goes.
00:00 – 00:33
01 · Cold open: coding is now ad-funded
Nick opens with a joke about coding becoming ad-funded like America, then introduces FreeBuff, a platform that swaps AI-agent subscription fees for in-app ads, explicitly disclaiming any sponsorship or affiliation.
00:33 – 01:42
02 · Free models and daily hour limits
A tour of FreeBuff's homepage shows five free models with different daily hour allowances, from 5 hours of GPT-5.6 Luna up to 20 hours of GLM 5.3 Flash, plus an intelligence/speed benchmark chart and a savings calculator claiming $11,400/year versus paid alternatives.
01:42 – 02:37
03 · Downloading and signing in, no credit card
Nick downloads the FreeBuff desktop app for macOS and signs in with Google, noting the app explicitly requires no credit card to start.
02:37 – 03:31
04 · How Freebucks and pricing work
An onboarding modal explains 'Freebucks': a daily allowance that refills at midnight Pacific with no session caps, spendable on any model, priced once per hour at session start; a separate paid-plan page shows $6.25-$46.67/month tiers for heavier use.
03:31 – 04:54
05 · Building a one-page bio site
Nick prompts GPT-5.6 Luna to build a one-page personal site about himself, watching the agent's live queue, file diffs, and an in-app ad (a dog food subscription box) surface at the bottom of the screen while it works.
04:54 – 06:05
06 · Building a stick-figure walking game
He opens a second parallel thread on GLM 5.3 Flash and prompts a simple stick-figure walking game, comparing its faster, rougher reasoning style against the slower Luna model running in the other tab.
06:05 – 07:03
07 · Missions, agent settings, and more ads
Nick reviews additional agent settings including a 'Mission' mode that drives a task to a merged pull request unattended, an AGENTS.md config option, and phone mirroring via a FreeBuff iOS app, while three ads are now visible across the two open build tabs.
07:03 – 08:38
08 · Reviewing both finished builds
The one-page bio site renders with a real headline, stats, and layout; the stick-figure game runs as a flat side-scroller where a stick figure simply walks, with two instances running at once.
08:38 – 09:41
09 · Why free access to AI agents matters
Nick closes with a reflection that people without money to spend on frontier AI subscriptions are now able to build real things, arguing this narrows the gap between people who can afford the smartest models and those who can't.
Lines worth screenshotting.
- FreeBuff replaces AI coding subscriptions with ads, undercutting OpenCode's $120/year, Codex's $240/year, and Cursor's $720/year with a $0/year free tier.
- The platform's top free-tier model, GLM 5.3 Flash, comes with 20 hours of daily use at no cost, more than any other model on the list.
- Signing up required no credit card at any point, a deliberate contrast to how most AI coding tools gate access behind payment info.
- FreeBuff's own comparison chart scores GLM 5.3 Flash at 45 on an intelligence index versus 53 for Fable 5.1 and GPT-6 Astro, meaning the free models trail frontier models by a real but not disqualifying margin.
- Freebucks refill to a fresh daily pool at midnight Pacific with no earning mechanic and no weekly or monthly session caps, so the entire cost model reduces to what you spend that day.
- One hour of any model, cheap or expensive, is priced once at the start of the session rather than metered continuously.
- FreeBuff can run two coding threads in parallel on different models at once, letting a user compare a faster/cheaper model against a slower/pricier one on the same kind of task side by side.
- A single, moderately detailed prompt produced a working, reasonably designed one-page bio site inside roughly a minute of visible agent 'thinking' time.
- In-app ads surface at the bottom of the build screen while the agent works, functioning like a loading-screen ad rather than a hard paywall interruption.
- Running two build threads simultaneously visibly spiked the presenter's laptop fan and heat output, a reminder that 'free' AI compute still has a local resource cost.
- FreeBuff also sells a paid tier ($6.25 to $46.67/month) for heavier use, meaning the free tier functions as an acquisition funnel rather than the entire business.
Ad-funded AI agents make coding free, not frontier-grade
WHAT TO LEARN
FreeBuff swaps a subscription for ad views and still produces working code, which matters most for people who were priced out of AI coding tools entirely, even though the free models trail frontier ones on capability.
01Cold open: coding is now ad-funded
- FreeBuff funds AI coding models through in-app ads instead of a subscription fee, which is worth knowing about if the sticker price on tools like Codex or Cursor is what's stopping you from trying agentic coding at all.
02Free models and daily hour limits
- Free daily hour allowances vary a lot by model: GPT-5.6 Luna gets 5 hours while GLM 5.3 Flash gets 20, so which model you pick changes how much free coding time you actually have.
- FreeBuff's own benchmark chart scores its free models below top-tier closed models like Fable 5.1 and GPT-6 Astro on intelligence, meaning free access here trades some capability for zero cost, not a like-for-like swap.
03Downloading and signing in, no credit card
- The entire signup flow required no credit card, just a Google sign-in, which removes the single biggest friction point that stops people from ever trying a coding agent.
04How Freebucks and pricing work
- Session pricing is charged once per hour at the start of a session rather than metered continuously, so cost is predictable before you commit to using a model.
- There's no weekly or monthly cap, no way to earn extra credits, and no rollover; the entire budget resets to the same daily pool at midnight Pacific, simplifying the mental math of how much you have left.
05Building a one-page bio site
- A single, moderately detailed prompt was enough for a free-tier model to produce a real bio page with a headline and stats section in a few minutes.
- In-app ads appeared at the bottom of the screen during generation, functioning like a loading-screen ad rather than a hard interruption; the cost of 'free' showed up as attention, not a blocked workflow.
06Building a stick-figure walking game
- Running two build threads on two different free models side by side is a fast, low-cost way to compare how a faster/cheaper model and a slower/pricier one handle the same category of task.
- The cheaper, faster-reasoning model visibly spiked local CPU and fan load when run in parallel with another session, a reminder that 'free' compute still has to run somewhere and can hit your own machine's limits.
07Missions, agent settings, and more ads
- A 'Mission' mode that autonomously completes a task, verifies it, commits, opens a pull request, and merges it once checks pass shows these free-tier tools are being built toward unattended, production-adjacent workflows, not just toy demos.
08Reviewing both finished builds
- The final outputs were functional but modest: a working personal site and a stick figure that mostly just walks, an honest baseline for what a single one-shot prompt on a free-tier model actually produces.
09Why free access to AI agents matters
- People who can't afford frontier-model pricing are functionally locked out of the productivity gains those models unlock, which compounds the gap between who can and can't build with AI.
- Free, ad-funded access to even mid-tier models is framed as a way to narrow that gap, even though the video is candid that the free models aren't as capable as the frontier ones.
Terms worth knowing.
- Freebucks
- FreeBuff's daily spending allowance for AI model usage. It refills every day at midnight Pacific with no rollover, no earning mechanic, and no weekly or monthly cap.
- Artificial Analysis Intelligence Index
- A third-party benchmark score used to rank AI models' relative capability; higher numbers indicate stronger benchmarked performance, shown as the axis on FreeBuff's model comparison chart.
- Mission mode
- A FreeBuff agent setting, described in-app as driving a coding request through to completion unattended: verified, committed, opened as a pull request, and merged once checks pass.
Things they pointed at.
00:18toolOpenCode
00:18toolCodex
00:18toolCursor
01:02toolGPT-5.6 Luna
01:02toolDeepSeek V4.1 Flash
01:02toolGLM 5.3 Flash
Lines you could clip.
00:00
“Well boys, coding is now freer than America.”
00:03
“just like america it's funded by ads”
06:12
“complete the user's request fully. when the work is correct, lean and verify. commit it. open a pull request. drive its checks to green and merge it.”
08:01
“we're now running two instances of stick walk, ladies and gentlemen... the game is just walking. we are just walking. wow.”
09:26
“us permanent underclassmen can at least start using models for free”
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.
analogy
Well boys, coding is now freer than America. and just like america it's funded by ads to make a long story short these guys freebuff made a platform where instead of pay for ai agent tokens using your credit card you do so by watching advertisements and so whereas open code might cost 120 bucks a year codex might cost 240 and so on and so forth freebuff does it all for zero bucks a year and in this video i'm going to actually walk through it and then do some ai agent coding to show you guys how it works and just before anybody gets in my face about being sponsored we're not affiliated whatsoever i just think think it's cool that people that don't have money can now use AI agents for free.
Just checking out the website here, they got a bunch of breakdowns of the models that you get access to. So five hours of GPT 5 .6 Luna for free, DeepSeek V4 .1 Flash for six hours, also quite comparable benchmark wise. MuseSpark 1 .2, which is a little bit dumber, but still pretty solid.
And then they have a couple of other ones, which I don't think anybody's ever heard of. And interestingly enough, GLM 5 .3 Flash, which I benched a while back is like pretty solid, is up to 20 hours. is now.
And I mean, I don't know if you guys have seen these charts, but the models that they're providing us like GLM 5 .3, which scores 45 on intelligence compared to, I don't know, Claude Fable 5 .1 is 53, or maybe DeepSeek V4 .1 Flash, which scored 40 to GPT -6 Astro, which scored 53. You know, these are actually pretty damn good models to be using, especially in coding capacities.
Because, you know, some of these models like 5 .6 Luna are available for free as like chat GPT, you know, chat bots. But this now allows you to build things with them, which I think is a major unlock. They also have a little like calculator here.
So you could see just how much you'd save just watching ads all day. Not actually watching ads all day, but you know, their ads is part of it. And yeah, they have a bunch of desktop apps and stuff like that.
So let's not waste our time. Let's actually download this for Mac OS. Here is the entire flow.
I'm just going to download this unverified file. Please, for the love of God, don't hack my computer. OK, applications it is.
And boom, looks like we're good to go. I'm now going to open up a free buff here. And they say that you don't need a credit card to do any of this, which sounds pretty interesting to me.
I guess all we need to do is just sign in. I'm just going to sign in with Google. And after the sign in, it'll return to my terminal.
It looks like there we go. So first thing that's kind of confusing is it says start for $5 right over here and you click it and you go over to freebuff .com, which is the app that you're using. I think what they're trying to do here is to show you what the ads that you're going to have at the bottom of your screen are going to look like.
That seems okay to me. It's not super disruptive or anything. And obviously if I'm in a country where it's hard to come by cash, like I'm not going to mind watching an ad or two if it gives me the ability to use something, right?
Anyway, why don't I just call this project because we name all of our project folders project. And here is the service. So sessions are now bought with free bucks.
It is a daily allowance you can spend on any model. There's a fresh pool every day. Your daily free bucks refill at midnight Pacific.
There's nothing to earn, nothing to wait for. No more weekly or monthly session caps. The only thing that counts is what you spend.
An hour of any model is one price charged once when the session starts. You can also spend it however you like. So pick the cheap one all day or save up your free bucks for the expensive ones.
That's kind of cool. Looks like we have it open right over here. Why don't we get started?
Let me see. So what are we clicking on down here? Looks like here are the various prices.
So $5 an hour, basically 20 hours per day of GLM 5 .3 Flash. That's insane. If I go down to GPT 5 .6 Luna, looks like we're spending 20 an hour, which is still like five, which seems totally cool to me.
So I think what I'm gonna do is I just really wanna see the ad flow. I'm just gonna click this. I'm gonna say, make me a one page site.
nick surayev look up what you need make me sound super cool but also don't go over the top design wise and i'll test this out on some other stuff later don't worry um okay so what am i seeing here the second i did it looks like there is now a little box on the right hand side showing what's in the queue shows the changes the diffs between befores and after shows some files a little preview and then even a terminal that we can communicate with as well so this is starting to look like a pretty fleshed out ide let's just full screen this because why the hell not looks like they also have the ability to go from build mode to plan mode which is cool so this is reminiscent of sort of like early clod fabling Okay.
And I think this is where the ads are going to come in. So get 50 % off your first box, fuel your dog's lifestyle with the farmer's dogs, ready to serve human grade meal plans. That's pretty interesting.
It looks like they're just listing the ads sort of where you are expecting, expected to look as somebody waiting for a token to be output. I just wanted to make it clear that like you guys could actually use this. So yeah, I mean, it's been like a minute in a minute and 10 seconds.
We're not seeing anything yet. We did see another ad swap. And just so that you guys know, my computer's starting to get a little hot, which is kind of scary.
Why the hell is my computer getting hot? That's odd. Ooh, it's very loud.
I think realistically, what we'll probably do is open multiple tabs. But do you see how it says one tab in use? Hmm, that's interesting.
Maybe you can't actually open multiple tabs. Well, why don't we give it a try? Maybe while I'm doing this, let's open a project in project.
Okay. Just wanted to like actually have two simultaneously. Okay, nice.
My threads are actually capable of talking to each other. And can I rename this? Let's call this two.
This up here will be one. This over here is going to be a, and I want to use a different model here to show you guys how that works. Going to click dismiss there.
Make me a super simple stick figure walking game. And then I'm just going to go enter. Now we're using a different model here.
This is GLM 5 .3. So I'm expecting the UX to look a little bit different. And that's what's going on here.
It's just thinking, thinking, and now it's red. So the reasoning on GLM 5 .3 Flash is going to be crazy fast relative to, you know, Luna. But obviously performance on Luna is going to be a lot better.
Speaking of Luna, it looks like we're still writing the index at HTML, which is kind of neat. If I go back here, you can see that it's now actually doing a bunch of calculations. That's GLM 5 .3 to, like, figure out my stick figure, which is pretty neat.
What does mission mean? Merge PR. Complete the user's request fully.
When the work is correct, lean and verify. Commit it. Open a pull request.
Drive its checks to green and merge it. That's interesting. I don't think I'm going to select that, but it's a very nice feature to have.
And then what else can I do? I could, like, feed in my agents .md. We got the ability to like set up a bunch of project settings here.
I could even mirror it to my phone via FreeBuff iOS. That's pretty neat. So I can actually do this coding for free while I'm out and about.
And it doesn't look like it's too bad here. We have three ads. We have the dog box one up here, runnable over here, and then another one over there.
So like I'm not exactly... Like, I'm not super pissed just yet, you know? I feel like you go way too far with this stuff, by the way.
Like, it's really easy to, like, go mad and, like, force people to watch, like, a five -minute Mountain Dew commercial before you take your first sip, but it doesn't look like they're at that point yet. Maybe one day. Honestly, the experience seems quite similar to just using most other agents.
That doesn't look super good. What the heck's going on over here? Make the future useful?
Hmm. I think we're gonna have to... upgrade the design here guys well this thing certainly ain't winning any awards for copywriting that's for sure you know all this puppy down here is you know solving the navier -stokes equation don't mind me just doing some serious agentic engineering okay there we go and it looks like that was the way so uh pretty cool to see the site obviously we also have the ability to select an element to add to a message so i can actually select a specific thing and then like give those instructions to The agent, if I just zoom out of this, I'm not going to pretend this is, you know, the most groundbreaking thing I've ever seen at this point.
I think it's neat that AI can whip up websites like this in 0 .2 seconds and all. But, you know, nothing really super crazy here, aside from the fact that it was totally free for me to use. I didn't even need to give them a credit card.
I didn't need a ChatGPT subscription at all. And yeah, I mean, that part's kind of neat. All right, let's take a look at our stick figure game.
I just said, let's run it, open it for me. And here's our little stick walk. Nice.
Not bad. Oh, okay. Here's another one.
We're now running two instances of stick walk, ladies and gentlemen. And that's it. The game is just walking.
We are just walking. Wow. How far could he walk, I guess?
100 meters? 200 meters? I wanted it to be fun.
I don't think it's really all that fun, but suffice to say, you guys can code stuff, make games and whatnot. And yeah, I wish I had something more interesting to show you there at the end, but hey, figured I might as well give you guys a... a video that is real as opposed to one that's like this free agent will blow your socks off.
It's, it's free, right? Like you guys just look at ads as you guys saw there. And, um, you know, I guess we're going to figure out the economics of that business model pretty soon.
But for now it seems, I think an issue with, you know, how agents have been distributed over the course of the last like four years or so is obviously, especially now frontier intelligence costs a lot of money. And so I'm just ripping, you know, fable 5 .1 and Astro all the time. And I'm capable of doing really cool things.
And then there are a bunch of people that are sending me messages like Nick, like he has not. really that good. And I'm like, no, dude, you got to use the new ones.
And they're like, what do you mean? Oh, that's super expensive. I can't afford it.
And it's like, okay, obviously people that have access preferentially to, you know, the smartest intelligence are going to be able to do cooler things with the intelligence than people that don't have access to it. So my hope is this in some way, shape or form, at least equalizes a little bit of that disparity. Obviously, you know, the super intelligent models are behind.
opaque, you know, error -gapped doors now with open AIs and Anthropix like super smart ones. But hey, us permanent underclassmen can at least start using models for free. Just wait till they strap a Neuralink to us and force us to watch ads all day.
Catch y 'all later.
The Hook
The bait, then the rug-pull.
The title promises a way to run Codex-style coding agents for nothing, and the opening line delivers the joke before the pitch: coding, like the country hosting it, now runs on advertising. What follows is a straight walkthrough of FreeBuff, an ad-funded agent platform, then two live builds to see if 'free' actually produces working code.
Named ideas worth stealing.
02:37concept
Freebucks daily allowance
FreeBuff's pricing model: a shared daily currency ('Freebucks') refills at midnight Pacific, spendable on any model, with each hour-long session priced once at start rather than metered per token, and no weekly/monthly caps.
Steal forany tool considering a usage-based free tier instead of a hard subscription paywall
How they asked for the click.
VERBAL ASK
00:18product
“we're not affiliated whatsoever, I just think it's cool that people that don't have money can now use AI agents for free”
Nick explicitly disclaims sponsorship on camera, even though the video's description promotes both a FreeBuff link and his own paid Maker School community, a disclosed non-affiliation paired with monetized description links.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
