Latest / AI Fire Daily / #111 Robin: This Free AI Coding Agent Gives You GPT 5.6 Luna for Hours - The Catch Is Ads

Transcript

#111 Robin: This Free AI Coding Agent Gives You GPT 5.6 Luna for Hours - The Catch Is Ads · AI Fire Daily

• 0:00
  Intro music AI coding tools are getting incredibly
• 0:02
  expensive. OpenCode is around $120 a year. Codex
• 0:07
  sits at $240. Beat. Now imagine getting full
• 0:12
  access to an AI coding agent for $0. You just
• 0:16
  pay with a little bit of your attention. Right,
• 0:18
  and that completely flips the traditional software
• 0:20
  model on its head. I mean, it's a massive shift.
• 0:23
  Welcome to the deep dive. Today we're exploring
• 0:26
  a new ad -supported tool called FreeBuff. we're
• 0:29
  going to unpack how its daily free bucks economy
• 0:32
  functions. Yeah, the economy part is fascinating.
• 0:34
  It really is. Then we'll explore its surprisingly
• 0:37
  robust agent workspace. And finally, we'll analyze
• 0:40
  some real -world tests to see if free AI can
• 0:43
  actually build functional software. I'm really
• 0:45
  excited about this one. This could totally change
• 0:47
  how beginners and hobbyists experiment with code.
• 0:50
  You get access to real development tools with
• 0:52
  absolutely zero financial risk. To understand
• 0:55
  the value here, we should probably start with
• 0:57
  the economics. We need to establish the baseline
• 0:59
  before we test the actual performance. Right,
• 1:02
  because OpenCode and Codex charge heavy annual
• 1:04
  fees. But FreeBuff costs absolutely nothing.
• 1:08
  There is no credit card required at any point.
• 1:11
  So how does that actually work? Well, instead
• 1:13
  of a monthly subscription, you operate on a daily
• 1:16
  allowance. The platform calls this virtual currency
• 1:19
  FreeBucks. Two secs silence. A daily allowance
• 1:23
  genuinely changes the psychology of coding. You
• 1:26
  aren't trapped in a rigid monthly contract. It
• 1:29
  really does alter your mindset. Your balance
• 1:31
  simply resets every single night at midnight
• 1:33
  Pacific time. There are no weird weekly caps,
• 1:37
  and there are no hidden monthly limits. You wake
• 1:39
  up to a fresh pool of computational resources
• 1:42
  every morning. I could see that being incredibly
• 1:44
  liberating for a weekend hacker. But I also wonder
• 1:46
  about the psychological pressure. Do you feel
• 1:49
  an anxious urge to use your remaining hours?
• 1:51
  That's a very fair point. It can definitely induce
• 1:54
  a sort of use it or lose it anxiety. You might
• 1:56
  feel pressured to start a project just to burn
• 1:58
  those credits. Right. So how does that virtual
• 2:00
  currency translate into actual minutes of coding?
• 2:03
  Different AI models burn this allowance at very
• 2:06
  different rates. GPT 5 .6 Luna is the heavier,
• 2:09
  more capable model. It costs roughly $20 per
• 2:13
  hour in this virtual currency. So that gives
• 2:16
  you about five hours of daily usage. Five hours
• 2:18
  of continuous generation is a solid coding session.
• 2:21
  What if you need more time? Then you switch to
• 2:24
  GLM 5 .3 Flash. That's the cheaper, faster alternative.
• 2:28
  It only costs about $5 an hour. Oh, wow. Yeah.
• 2:31
  That stretches your daily allowance out to a
• 2:34
  massive 20 hours. You can essentially leave it
• 2:36
  running all day long. Are those the only two
• 2:39
  options on the platform? Not at all. You also
• 2:41
  have access to DeepSeq V4 .1 Flash. That gives
• 2:45
  you around six hours of usage. OK. And then there's
• 2:48
  MuSpark 1 .2, which offers several hours as well.
• 2:51
  But the real discussion here is about the capability
• 2:53
  gap. Let's define our terms for the listener.
• 2:55
  When we talk about frontier models, BEAT, the
• 2:59
  most advanced capable AI systems currently available
• 3:01
  today. Perfect definition. We're talking about
• 3:03
  the big premium names. Claude Fabel 5 .1 and
• 3:07
  GPT -6 Astra are frontier models. Those systems
• 3:10
  score around a 53 on standard intelligence benchmarks.
• 3:14
  And where do these free ad -supported models
• 3:16
  sit on that exact same scale? GLM 5 .3 scores
• 3:19
  around a 45. DeepSeq v4 .1 flash sits near 40.
• 3:24
  So there's a clear mathematical gap in sheer
• 3:26
  reasoning power. Exactly. These are not the absolute
• 3:29
  bleeding edge systems. Two sec silence. You know,
• 3:33
  I still wrestle with prompt drift myself. Oh,
• 3:35
  yeah. When an AI loses context over a long conversation,
• 3:39
  it forgets earlier instructions. That makes me
• 3:42
  wonder about this capability gap. How so? Maybe
• 3:45
  having more hours with a slightly less intelligent
• 3:47
  model is actually beneficial. You can tinker
• 3:49
  more. You can fix everyday coding mistakes without
• 3:52
  stressing about a ticking clock. That's a really
• 3:54
  profound way to look at it. A benchmark score
• 3:56
  of 40 is plenty for basic UI work. Right. You
• 3:59
  don't need a genius neural network to change
• 4:01
  a button color. You just need patience and rapid
• 4:04
  iteration. Which leads me to a really practical
• 4:06
  question for you. How does the gap between a
• 4:08
  45 benchmark and a 53 benchmark actually feel
• 4:11
  when you're just trying to fix a basic UI bug?
• 4:14
  It's the difference between simple execution
• 4:16
  and deep architectural reasoning. A 45 can fix
• 4:20
  a typo in your CSS perfectly, but a 53 will look
• 4:24
  at that typo, analyze your entire rendering logic,
• 4:28
  and rewrite it for better performance. For basic
• 4:31
  bugs, you honestly won't notice the missing eight
• 4:34
  points. So lower scores just mean focusing on
• 4:37
  simpler everyday tasks. Exactly. It's simply
• 4:40
  about matching the specific tool to the specific
• 4:43
  job. So if you're constantly managing this finite
• 4:46
  daily budget, Every minute counts. You can't
• 4:49
  afford to fight a clunky interface. No, you really
• 4:51
  can't. How does FreeBuff's actual workspace help
• 4:54
  you maintain your flow state? Well, the setup
• 4:56
  process is surprisingly frictionless. You just
• 4:58
  download a native macOS application, then you
• 5:01
  sign in with your Google account. Okay, so no
• 5:03
  technical hurdles, though... Looking at the notes,
• 5:06
  there was one confusing detail during setup.
• 5:08
  Oh, yeah. They have this weird start for $5 button
• 5:11
  on the home page. Right. It honestly feels like
• 5:13
  a user interface mistake. Clicking it just takes
• 5:15
  you to a demo site. It simply shows you what
• 5:17
  the ad -supported experience looks like. That's
• 5:19
  confusing. It is. But once you bypass that, the
• 5:22
  native interface is incredibly clean. We should
• 5:25
  probably clarify a crucial distinction right
• 5:27
  now. Beat. What's the exact difference between
• 5:30
  a standard chatbot and a true coding agent? A
• 5:33
  chatbot just... spits raw text onto your screen,
• 5:36
  you have to manually copy and paste that code
• 5:38
  into your editor. Right, which is tedious. Exactly.
• 5:41
  A true coding agent actually works across your
• 5:44
  entire local project. It creates new files. It
• 5:47
  directly edits your existing code. It actually
• 5:49
  runs terminal commands too, right? Yes. And it
• 5:51
  literally inspects its own changes. It has agency
• 5:54
  within the actual file system. Exactly. It's
• 5:57
  less like talking to a standard chatbot. It's
• 5:59
  more like pair programming with a junior developer.
• 6:02
  They have full control of your keyboard. You
• 6:04
  point at the screen, tell them what to fix, and
• 6:06
  watch the code change. That's a massive leap
• 6:08
  in utility. So what does that environment actually
• 6:11
  look like? The interface feels exactly like a
• 6:14
  lightweight modern IDE. You have a chat box for
• 6:17
  your instructions on the left. Then you have
• 6:19
  a right side panel that houses all of your active
• 6:22
  tools. Usually when I use web -based AI, I spend
• 6:25
  half my time wrestling with visual disconnects.
• 6:28
  I'm assuming a native app solves that workflow
• 6:31
  problem. Completely. In that tool panel, you
• 6:33
  get a live task queue. You see real -time file
• 6:36
  changes and visual diffs. There's a live preview
• 6:40
  window. You even get an integrated terminal.
• 6:43
  Plus, there are dedicated build and plan modes
• 6:45
  for structuring complex apps. It sounds incredibly
• 6:48
  cohesive. You just watch the logic build dynamically
• 6:50
  in real time. It's like stacking Lego blocks
• 6:53
  of data. That's a great analogy. But the advanced
• 6:55
  features are what really caught my attention
• 6:57
  here. Mission mode seems particularly ambitious
• 7:01
  for a completely free tool. Mission mode is honestly
• 7:05
  wild. You give the agent a complex multi -step
• 7:08
  request. And it handles it. It doesn't just write
• 7:10
  a block of code. It literally opens a pull request.
• 7:13
  Wait, it handles version control automatically?
• 7:15
  Yes. And because it has terminal access, it can
• 7:18
  run your automated test suite. That's crazy!
• 7:21
  If a test fails, it reads the red error logs.
• 7:24
  It feeds that error text back into its own context
• 7:27
  window. Then it rewrites its own code until the
• 7:29
  test passes. Finally, it merges the PR. That's
• 7:33
  a complete automated software development lifecycle.
• 7:36
  Beat. I also read that it supports custom instruction
• 7:38
  files. Right. You can load an agents .md file
• 7:41
  directly into your project root. This tells the
• 7:45
  agent exactly how to behave in your specific
• 7:47
  code base. You can dictate strict style guides.
• 7:50
  So it doesn't just write generic code? Exactly.
• 7:52
  You can enforce architectural rules so it doesn't
• 7:55
  break your existing patterns. And the precision
• 7:57
  of the UI selection tool? is fascinating. It
• 8:00
  really is. Instead of typing out a long, vague
• 8:02
  prompt, you can just point. Like visually. Yeah.
• 8:05
  You select a specific UI element right inside
• 8:07
  the live preview window. You send that exact
• 8:10
  DOM element to the agent and say, fix this specific
• 8:14
  button's padding. That eliminates so much conversational
• 8:17
  ambiguity. Two -second silence. Whoa. Imagine
• 8:21
  mirroring a complex build right to your phone
• 8:23
  while getting coffee. Right, and that's an actual
• 8:25
  feature they shipped. The FreeBuff iOS app allows
• 8:28
  for complete session mirroring. No way! Yeah,
• 8:32
  you can literally walk away from your laptop.
• 8:34
  You can monitor the agent's progress in the terminal
• 8:37
  directly from your iPhone. That feels like a
• 8:39
  very real glimpse into the future of decentralized
• 8:42
  work. I also noticed the parallel processing
• 8:45
  capabilities in the test notes. This part is
• 8:48
  slightly confusing in the actual UI. It constantly
• 8:51
  displays a message saying one tab in use, but
• 8:54
  you can actually run multiple tabs simultaneously.
• 8:57
  You mean using different AI models at the exact
• 8:59
  same time? Yes. You could have Luna building
• 9:02
  a complex landing page in tab one. Meanwhile,
• 9:05
  GLM is rapidly testing a game mechanic in Tab
• 9:08
  2. Which naturally makes me wonder about the
• 9:10
  underlying economics of that setup. Does running
• 9:12
  parallel projects with different models drain
• 9:15
  the single free bucks pool twice as fast? It
• 9:17
  absolutely does. You're drawing from the exact
• 9:19
  same finite daily reservoir. I see. Running two
• 9:22
  heavy models simultaneously will deplete your
• 9:25
  daily hours incredibly quickly. You are essentially
• 9:27
  burning the candle at both ends. Running parallel
• 9:30
  tabs drains your daily allowance from both ends.
• 9:32
  Exactly. You have to manage those compute resources
• 9:34
  very carefully. Now, balancing those resources
• 9:37
  is tricky. A great user interface and free computational
• 9:41
  hours sound perfect on paper. But what is the
• 9:45
  actual cost to your attention? Right. That's
• 9:47
  the big question. And more importantly, can this
• 9:50
  specific tool actually write good code? Well,
• 9:53
  the catch is obviously the advertising model.
• 9:55
  But the ad experience is surprisingly mild. You
• 9:59
  see roughly three small banner ads on the screen
• 10:02
  at any given time. Where are they placed? Do
• 10:04
  they interrupt the actual coding workflow? Not
• 10:06
  at all. They sit in completely normal interface
• 10:08
  spots. Like where? You might see an ad for the
• 10:11
  farmer's dog, or maybe a banner for Runnable,
• 10:14
  but they just rotate quietly in the background.
• 10:16
  OK, good. They don't block your code or your
• 10:19
  terminal. I have a measured concern about this
• 10:22
  trajectory, though. Have you been? If they shift
• 10:25
  to force pre -roll videos later, the entire workflow
• 10:29
  breaks. Oh, absolutely. A 15 -second unskippable
• 10:32
  video would destroy a coder's flow state completely.
• 10:35
  That's the constant looming risk with any ad
• 10:38
  -supported software. The monetization strategy
• 10:41
  can always pivot. Right. But right now it's perfectly
• 10:43
  fine. It just blends into the background of the
• 10:45
  IDE. Let's dig into the actual coding performance.
• 10:48
  You ran two specific real -world tests to see
• 10:52
  what these models can actually do. Yeah, test
• 10:55
  one utilized GPT 5 .6 Luna. The goal was building
• 10:59
  a simple responsive one -page personal website.
• 11:02
  A very standard well -documented baseline test.
• 11:06
  How did Luna handle the initial request? It was
• 11:09
  noticeably sluggish. There was a distinctly long
• 11:12
  delay before it even started typing out the architectural
• 11:15
  plan. Interestingly, my laptop actually started
• 11:18
  getting hot and loud while it processed the task.
• 11:21
  That suggests some heavy local resource utilization.
• 11:24
  Yeah. Or perhaps it's just a highly demanding
• 11:26
  electron app bridging with the cloud. Could be.
• 11:28
  What about the final visual output? The site
• 11:30
  worked. The basic HTML structure was entirely
• 11:32
  there. The agent successfully created a usable,
• 11:36
  functional web page. OK, so the header and footer
• 11:38
  sections were all laid out properly. Yes. But
• 11:40
  there were obvious drawbacks to the final output,
• 11:43
  right? Yes. The design itself was incredibly
• 11:45
  plain. It completely lacked any real aesthetic
• 11:49
  polish or modern styling. And what about the
• 11:51
  generated copy? The text definitely needed a
• 11:55
  heavy human touch. It felt very robotic and generic.
• 11:58
  I would have to run another full round of prompts
• 12:01
  just to polish it. I'd need to ask it to improve
• 12:04
  the CSS layout and completely rewrite the placeholder
• 12:06
  text. So Luna is clearly better for tasks where
• 12:10
  structural quality trumps raw speed. Exactly.
• 12:13
  You spend more of your free bucks for a slightly
• 12:15
  better architectural foundation. Yeah. But you
• 12:18
  actively trade off generation speed and local
• 12:20
  system resources to get it. Test two is quite
• 12:22
  different in scope. You switched over to the
• 12:25
  lighter GLM 5 .3 flash model. Yes. For this one,
• 12:28
  I requested a stick figure walking game. Oh,
• 12:30
  fun. I really want to to see how a lighter model
• 12:33
  handled basic interactive logic and canvas rendering.
• 12:36
  How was the reasoning speed compared to the Luna
• 12:38
  test? The difference was obvious almost immediately.
• 12:41
  GLM reasoned incredibly fast. Yeah, it started
• 12:45
  generating the JavaScript right away without
• 12:47
  that long initial delay. And the final playable
• 12:50
  output of the game? It was incredibly basic.
• 12:53
  The game technically executed without crashing.
• 12:56
  OK. You had a crude stick figure and it walked
• 12:59
  across the screen. But that was the entire interactive
• 13:01
  experience. It lacked any real depth or physics.
• 13:05
  Were there any technical hiccups during the actual
• 13:08
  generation process? It glitched out badly at
• 13:11
  one point. It accidentally launched two instances
• 13:14
  of the game simultaneously. Why did it do that?
• 13:17
  It likely got a little confused by its own preview
• 13:19
  window reading the DOM. Ah, I see. The preview
• 13:22
  hot -reloaded the page while the agent was still
• 13:25
  injecting a new canvas element. It basically
• 13:27
  tripped over its own shoelaces. That sounds like
• 13:30
  the classic development trade -off. Yeah. GLM
• 13:32
  is perfect for quick prototypes where you don't
• 13:34
  want to burn your allowance. Exactly. Fast, cheap,
• 13:38
  and just good enough for simple interactive work.
• 13:40
  Right. For basic features or small layout experiments,
• 13:44
  you should definitely save Luna for the heavier
• 13:46
  architectural lifting. Based on these real -world
• 13:48
  tests, I have to ask a blunt question. Is FreeBuff
• 13:52
  actually a legitimate threat to premium paid
• 13:56
  coding agents right now? Honestly, not right
• 13:58
  now. It absolutely will not replace a $20 monthly
• 14:02
  subscription for professional full -time developers.
• 14:06
  Because of the polish. Yeah, the outputs just
• 14:08
  aren't polished enough yet. But it is an incredible
• 14:11
  zero -risk sandbox for literally everyone else.
• 14:15
  It won't replace premium tools, but it's a perfect
• 14:17
  free sandbox. That's the perfect way to frame
• 14:19
  it. It's a phenomenal learning environment. We
• 14:22
  need to pull all these technical threads together.
• 14:24
  Why does this ad -supported model actually matter
• 14:26
  for you, the listener? Maybe. FreeBuff's real
• 14:29
  value is clearly not flawless immediate code
• 14:31
  generation. No, the generated code still needs
• 14:34
  a lot of active human guidance. The real underlying
• 14:37
  value here is democratizing the entire workflow
• 14:40
  of AI agents. It teaches you how to structure
• 14:43
  a prompt. It teaches you how to guide an autonomous
• 14:46
  AI through a multi -step project. It gives you
• 14:49
  a zero -dollar entry fee to the future of coding.
• 14:51
  You can learn how to actually manage these digital
• 14:53
  entities. You can test small ideas instantly.
• 14:56
  without opening your wallet. And crucially, you
• 15:00
  avoid that creeping subscription fatigue. Oh,
• 15:02
  absolutely. You don't have to internally justify
• 15:05
  another recurring monthly bill just to prototype
• 15:08
  a silly weekend project. You simply log in, watch
• 15:12
  a few banner ads, and start building. It lowers
• 15:14
  the barrier to entry to absolute zero. That fundamentally
• 15:18
  changes the demographics of who gets to participate
• 15:21
  in building software. It totally shifts the power
• 15:24
  dynamic. Anyone with a good idea and a little
• 15:26
  bit of patience can now spin up a functional
• 15:28
  prototype. Yeah. That's a beautiful thing for
• 15:31
  global creativity. I want to leave you with a
• 15:33
  final thought today. Two -second silence. If
• 15:36
  global compute costs keep rising, what happens
• 15:39
  when our most powerful thinking tools are all
• 15:42
  completely subsidized by advertisers? Will the
• 15:44
  AI of the future subtly optimize its own code
• 15:47
  to keep us looking at banner ads just a little
• 15:50
  bit longer? That is a beautifully terrifying
• 15:52
  question. Thank you all for exploring us with
• 15:54
  us and keep building. Thank you for joining The
• 15:55
  Deep Dive. Out T -Row Music.