Transcript

00:00:00That little company known as Anthropic,
00:00:02of course, Claude Code, their famous tool behind it
00:00:05has caused quite the ruckus, okay?
00:00:07We're talking about people are very, very upset, okay?
00:00:11I swear I could still hear Twitter people
00:00:13just screaming right now and it's been days.
00:00:16And if you have no idea what I'm talking about,
00:00:18honestly, you're living a blessed life, okay?
00:00:20I want you just to look out on that farm
00:00:23and I just want you to smile
00:00:25because you have it peaceful, you have a nice time,
00:00:28but the rest of us have to live with what happens
00:00:30out here in the real world, okay, punk?
00:00:32You don't know what I have to see,
00:00:34but since you're watching this video,
00:00:35you will now discover everything, okay?
00:00:36We're gonna go over everything.
00:00:38There's a lot of people yapping
00:00:39about what's going on with Anthropic.
00:00:41I'm gonna give you hopefully what is the full story,
00:00:44but more, I'm gonna also just, I'm gonna make a couple,
00:00:47you know, I might project a little bit.
00:00:48I might make some guesses myself at the end, okay?
00:00:50For those that are in the know,
00:00:52you've probably seen the image.
00:00:53For those that aren't, you have not seen this image,
00:00:55but effectively it says this credential is only authorized
00:00:58for the use with Clod code and cannot be used
00:01:01for other API requests.
00:01:03This is the big image that kind of kicked off
00:01:05this whole Anthropic hate that's going on right now.
00:01:08So what exactly is happening here?
00:01:10Well, if you don't know, Anthropic does have plans.
00:01:12You can get a pro plan, a pro 5X plan, and a pro 20X plan.
00:01:16And if you're on that plan, you at least up until now
00:01:20could use that token and say cursor,
00:01:22you could use it inside of open code
00:01:24and be able to use your plan and use your set amount
00:01:27or limited tokens to be able to do a bunch of code changes.
00:01:30Well, that all changed this week.
00:01:32They said, hey, guess what?
00:01:34Absolutely not.
00:01:35Nobody else is doing that.
00:01:37You can only use the plan in Clod code.
00:01:39Now this came as a big surprise for everybody, okay?
00:01:41People were not happy.
00:01:43The GitHubs are just, as you can see,
00:01:45just filled with just people trying to find workarounds
00:01:47with it, tons of comments, ridiculous amounts of comments
00:01:50going on in there.
00:01:51People saying they're canceling, just a lot of stuff.
00:01:54Lot of people, not happy.
00:01:56So in other words, if you pay for a subscription plan,
00:01:59you cannot use third-party tools.
00:02:01You must use Clod code.
00:02:03If you pay by the API pricing,
00:02:05which is literally like something like 10X more expensive,
00:02:09then you can use things that are non-Clod code.
00:02:11You can use open code, you can use cursor,
00:02:13you can use whatever, but there's a lot more to it.
00:02:15I think this Reddit post right here
00:02:17is honestly really illuminating.
00:02:19They're effectively saying that everybody's kind of lying
00:02:21about the situation, okay?
00:02:23People aren't giving the full truth,
00:02:25or at least they're giving maybe a partial truth
00:02:27is probably better.
00:02:28I mean, they are saying,
00:02:29"Hey, yeah, there's misinformation going on."
00:02:30Well, it's more like partial information.
00:02:32It turns out part of Clod code's big plans
00:02:36is they actually explicitly state in their TOS,
00:02:39"This is only use in our products."
00:02:42And so has this always been known?
00:02:44Yes, apparently this have always been known.
00:02:46Effectively, what a bunch of these tools do
00:02:48is that they use your OAuth token
00:02:50and they forge requests with specific headers and body.
00:02:53It's very fragile.
00:02:54Like if you've ever done this before,
00:02:56have you ever had to treat like a black box service
00:02:58and you have to kind of guess how it needs to behave,
00:03:01and then you get it up and running,
00:03:02and then they make a change
00:03:03and it just destroys your service
00:03:06until you come back up with the new change that they've made.
00:03:08It's always very fragile, super frustrating,
00:03:11but this is how all the tools work.
00:03:12And this TOS thing, it's been around since
00:03:15all the way back when Clod code released in early 2025.
00:03:18They have always strongly discouraged
00:03:20using their APIs this way.
00:03:22They put real effort into preventing it.
00:03:24So in other words,
00:03:25this problem that everybody's very angry about
00:03:27was actually just really a ticking time bomb underneath
00:03:30because Anthropic didn't enforce it.
00:03:33They allowed it to continue to fester for a long time
00:03:36and a bunch of people switching over.
00:03:38Don't worry, we're about though,
00:03:39like literally we're about to get the tinfoil out.
00:03:42We're just not, we're not that good.
00:03:43I want to make sure you guys really understand
00:03:45what's happening before I go off and yap project.
00:03:48Lastly, apparently the person running Clod code
00:03:50has came out and said this on Twitter,
00:03:52which said this, yesterday we tightened our safeguards
00:03:54against spoofing the Clod code harness
00:03:56after accounts were banned for triggering abuse filters
00:03:59from third-party harnesses using Clod subscriptions.
00:04:01Third-party harnesses using Clod subscriptions
00:04:03create problems for users
00:04:05and are prohibited by our terms of service.
00:04:07They generate unusual traffic patterns
00:04:09without any of the usual telemetry
00:04:10that Clod code harnesses provide,
00:04:12making it really hard for us to help debug
00:04:14when they have questions about rate limit usage
00:04:16or account bans,
00:04:17and they don't have any other avenue for the support.
00:04:20In other words,
00:04:21it's just that these abusive third-party platforms
00:04:24are just impossible to debug
00:04:26and any problem is perceived to be an anthropic problem,
00:04:29not the tools problem.
00:04:30And so that's what they're really saying.
00:04:32They're really just saying, hey, for your best interest
00:04:35and for our ability to debug,
00:04:37the only possible way forward is for us to make sure
00:04:40that we're always enforcing our TOS no matter what.
00:04:44All right, so that's kind of like the, what has happened?
00:04:47All right, everybody, we are now on the same page.
00:04:50Now, the real question is why did this happen and why now?
00:04:54Like the open code has been using anthropic for a long time.
00:04:58Cursor has a lot of these tools,
00:05:00apparently like cloud something or another
00:05:02that some other service is also using it.
00:05:05Like this has been around
00:05:06for months upon months upon months.
00:05:08So why enforce it now?
00:05:10Especially at the height of Clod code glazing.
00:05:15The first thing people kind of talk about,
00:05:16which honestly I think is the weakest
00:05:19and probably the most incorrect of them all is the cost.
00:05:23So remember there's three potential plans.
00:05:26There is going to be a pro plan
00:05:28and then you have the five X plan
00:05:31and then you have the 20 X plan.
00:05:33This one's approximately $20 a month.
00:05:35This one's a hundred, I believe, and this one is 200.
00:05:39And so often what you see people kind of citing is that,
00:05:42okay, people don't really make usage of the pro plan
00:05:46except for maybe just a little bit actually cost more.
00:05:49And then the, of the five X,
00:05:50maybe half the people cost more than the actual thing.
00:05:53And then for the 20 X,
00:05:54the large majority of them end up costing a lot more.
00:05:57And to what extent they cost, I don't know.
00:06:00So the general calculus, everyone runs on the math to like,
00:06:03oh, this is why they're doing it.
00:06:05Goes something along the lines of this, right?
00:06:06The integral of P over whatever its bounds is.
00:06:08I'm not going to write its bounds.
00:06:10Plus of course, the integral of the five X plan,
00:06:13plus the integral of the, what's it called, of the 20 X plan
00:06:17is going to be like minus the integral.
00:06:19Let's just call it whatever this line is.
00:06:20I'm just going to call it G of X
00:06:22'cause I have no idea what the hell it is.
00:06:24Let's just call it G of X.
00:06:25This thing has to be greater than zero
00:06:27and that's why they're doing it, right?
00:06:29The area underneath this curve is bigger
00:06:32than the area underneath these little price points,
00:06:35step ladder curves, right?
00:06:36Like that's what everybody's saying.
00:06:38Everyone's saying, okay, this is the real cost.
00:06:40That's that.
00:06:41Honestly, I don't think this is the problem.
00:06:43I mean, yes, I think money's the problem,
00:06:45but I don't think this particular thing
00:06:47is the motivating factor or why we're seeing it now.
00:06:50So to kind of put us in the right direction,
00:06:52I believe that DAX actually has a really good,
00:06:55like at least directionally correct tweet, which is okay.
00:06:59It's their business to have the right
00:07:00to enforce their terms however they like, which is true.
00:07:02Like they already said from the beginning,
00:07:03they don't want anyone but Claude Code
00:07:05using their subscriptions.
00:07:06So it technically is their right to be able to do this.
00:07:09They're not obligated to give anybody access
00:07:11to their services in any other way
00:07:13than they have set forth in their terms of service.
00:07:15And if you've paid for it and you are now upset about it,
00:07:18you don't have to continue to pay, just stop paying.
00:07:20So DAX's projection is that models aren't sticky enough.
00:07:25Yes, Opus is good, but so is Gippity5 too.
00:07:27There's things that Gippity5 too can do better
00:07:29and there's things that Opus does better.
00:07:31So it's not like the models are actually that sticky.
00:07:33What they really want is that they want their entire tooling,
00:07:36their entire stack to be what you use.
00:07:39When you use Claude, you use Claude Code.
00:07:43This way they own everything.
00:07:45They show that they have more value effectively.
00:07:48And I think this is a big thing,
00:07:49especially when it comes to Open Code,
00:07:51because you may not realize this,
00:07:52but Open Code has been growing wildly.
00:07:54They're about to hit one million monthly active users
00:07:58on their open source project,
00:08:00which means before it was just a toy project,
00:08:02but now that it's just shooting up so wild.
00:08:04And the reason, by the way, it's shooting up so wild,
00:08:06if you have not used Claude Code,
00:08:08like I'm not trying to be mean, Claude,
00:08:10but I feel like I kind of have to be mean
00:08:13because can we all just agree?
00:08:14Anthropic is the number one fear-mongering,
00:08:17hey, software engineering is gonna end company of them all.
00:08:21And Claude Code, the product,
00:08:23is just significantly far behind Open Code.
00:08:27They've been having screen flickering,
00:08:29just like it is like a dance club in your 2E editor
00:08:33that is honestly just like,
00:08:35I'm gonna get epilepsy from even looking at it.
00:08:37They've had this for months now.
00:08:39They tweeted about how they fixed 85% of it,
00:08:43which is an absurd fix.
00:08:45That's not a fix.
00:08:46You don't know what the hell's going on
00:08:47if you can only fix 85% of it.
00:08:49And second, how did you even come up with that number?
00:08:52And third, one day later, they had to roll back that change
00:08:55because other bugs were cropping up.
00:08:57What it shows to me is that the company
00:08:59in which is telling you that software engineering is over
00:09:01can't even run their own, like their own software.
00:09:04They're not making the changes
00:09:06that are needed to be competitive.
00:09:07And they're losing out to a bald man
00:09:10and a vegan from the Ozarks.
00:09:12It's just diabolical, these two.
00:09:14Yeah, by the way, everybody knows,
00:09:16you don't start a land war in Asia
00:09:18and you never go head to head
00:09:20when open source is on the line with a vegan, okay?
00:09:22Everybody knows that.
00:09:24And by the way, even with Claude Code,
00:09:25besides for the flickering,
00:09:26when you type in ultra think it's like this rainbow,
00:09:28I had it more than once shift off
00:09:31and have like the antsy escape sequences
00:09:33just being printed in plain old, regular text.
00:09:35I mean, just egregious bug after egregious bug.
00:09:38I couldn't expand code with control O
00:09:41while in a menu to accept the code changes.
00:09:44So it's like, okay, accept these code changes, yes or no?
00:09:46Well, if you want to see the code, expand it.
00:09:48Well, you can't leave the menu and you're just like,
00:09:50oh my gosh, please, what are you doing this?
00:09:51Claude, stop it, stop it, Claude, just let me, oh, I'm gonna.
00:09:55The tool is not good, okay?
00:09:57First off, an impressive software project
00:10:00and they've done good so far,
00:10:02but it is not that usable comparatively to open code.
00:10:04Open code is a beauty, okay?
00:10:06It's absolute beauty.
00:10:07It gives you all the information you know, right up front.
00:10:09It's actually well-designed.
00:10:10It's by developers who like to code
00:10:13for developers who like to code and it's obvious.
00:10:15So to me, when I see this graph
00:10:17and people talking about this,
00:10:18I don't think so.
00:10:23I don't think that's it at all.
00:10:24All right, people, we're gonna need to do something.
00:10:26I'm gonna need the tinfoil hat up here for a second, okay?
00:10:35All right, so this is what I think the real problem is
00:10:38that Anthropic is actually having, okay?
00:10:41I don't think it's just simply that previous graph.
00:10:44What I think it is,
00:10:45it is partially that previous graph right here,
00:10:47but I think there's a lot of lies in that previous graph, okay?
00:10:50I think the amount that they're subsidizing the Pro,
00:10:55the 5X and the 20X is much more
00:10:57than this graph leads you to believe.
00:10:59I think the cost that they actually have to run everything
00:11:02is significantly higher.
00:11:03First off, you gotta remember they have to do
00:11:05these huge amount of training runs that involve buying
00:11:07and being able to utilize a whole bunch of hardware
00:11:11and energy, which is just gonna be a gigantic number.
00:11:14Then you have the whole lifetime problem of the model.
00:11:16How, I mean, who here's out there still using 3.5
00:11:20when you got 3.7.4 and 4.5 on the table right now?
00:11:23You also have Gippity 5, Gippity 5.1, Gippity 5.2.
00:11:27You have all these other options that are better than 3.5,
00:11:30get 3.5, costed a huge amount of money.
00:11:34And so the lifetime like operation of these models,
00:11:38especially right now in such great change is rather low,
00:11:41but the cost to actually run it is rather high.
00:11:43Plus you have employees which are extremely expensive.
00:11:46The stock grants that some of these higher paid jobs
00:11:49are just getting paid outrageous amounts of money.
00:11:52Plus you also have to have all the infrastructure
00:11:54and all that crap being set up to be able to run
00:11:57all this stuff that involves all the continual running
00:12:00of energy, replacing GPUs and all that.
00:12:02And then on top of it, you got things like Nvidia
00:12:05just dropping brand new AI GPUs called Ruben
00:12:08that apparently is gonna deliver 10X in reduction
00:12:11in inference token costs and 4X reduction
00:12:14in number of GPUs to train MOE models.
00:12:17That means MOE models, okay?
00:12:19When you want MOE models, you gotta get the new chip.
00:12:22And so this is like this constant rat race
00:12:24of trying to stay ahead of all the competitors
00:12:26like Chat, Gippity, Google, XAI.
00:12:30I mean, let's just say, I know XAI really is only good
00:12:33at putting people in bikinis
00:12:34when they don't wanna be put in bikinis,
00:12:36but one day you never know, okay?
00:12:38You just never know.
00:12:40So there's a real subsidization problem going on here.
00:12:44These tokens aren't just subsidized by like
00:12:47just the pro users versus the 20X users.
00:12:50They're not subsidized by like 80%.
00:12:52I would not be surprised to find out
00:12:54that these things are like 98% being subsidized right now.
00:12:58Meaning the true cost of the max plan currently
00:13:02is 50X more than they're actually saying
00:13:04because they're still trying to recuperate
00:13:07all the other things beforehand.
00:13:09And then to remain competitive in the future,
00:13:11there is a continual nonstop march
00:13:14into every single latest piece of hardware
00:13:17and everything that they must put in there.
00:13:19RAM has already been purchased for all of 2026.
00:13:22So what they have is what they get.
00:13:25And it's astronomically expensive
00:13:27with Samsung being like,
00:13:28yeah, we're gonna just forex our price.
00:13:30You're gonna pay for it?
00:13:31And they're like, yes, absolutely, please, yes, sir.
00:13:34When would you like me to bend over and how often, right?
00:13:36Like they're just absolutely at the whims of everybody.
00:13:39And so if they can make a stack that's so competitive
00:13:43and then force people into this highly subsidized plan
00:13:47that can only be used on their stack,
00:13:49their stack becomes the sticking point, right?
00:13:51Because you can't use the cheap effective access
00:13:55to unlimited tokens.
00:13:56And then on top of that,
00:13:57you don't have access to other models
00:13:59when you don't want to use them, right?
00:14:01That's why OpenCode is so dangerous.
00:14:02It isn't because you can use your Mac subscription.
00:14:05It's because when I go, huh, let me try Jipity2.
00:14:08Oh, wow, Jipity52 did really good right there.
00:14:11Maybe I don't want Opus.
00:14:13All of a sudden, yikes.
00:14:15Anyway, so there you go.
00:14:16That's my thoughts on the whole thing.
00:14:18I know maybe I'm a little bit more
00:14:19of a fringe conspiracy theorist when it comes to this stuff,
00:14:23but I think that I personally see trouble in the water, okay?
00:14:26I think that that's why they're doing it
00:14:27because why would you torpedo the single highest amount
00:14:32of developer gratitude and nonstop Twitter glazing?
00:14:36It's so over.
00:14:37We're never coming back, Software Engineering.
00:14:40Opus is right.
00:14:41I mean, you've seen videos on saying Dario was correct
00:14:44that 90% of code is gonna be written in three to six months.
00:14:47I mean, it's just, it's fever pitched.
00:14:49We're talking about absolute fever pitched zealotry
00:14:53going out there on Twitter.
00:14:54And then they just take,
00:14:56they literally just take old yellow out and then bam!
00:15:00It is just insane.
00:15:02So there you go.
00:15:03That's it.
00:15:04That's all I have to say.
00:15:04Hey, I'm gonna try something a little unusual.
00:15:07You wanna press like?
00:15:07You wanna make a comment?
00:15:08Tell me where you think I'm wrong.
00:15:09Tell me where you think I'm right.
00:15:10I never usually invite these kinds of things
00:15:12'cause honestly I think it's kind of,
00:15:13I think it's kind of weird to do that.
00:15:15Yeah, I know I did push for subscribers
00:15:16for a little bit to get over a million.
00:15:18By the way, thank you for the million.
00:15:19I appreciate that.
00:15:19Anyways, thanks for watching.
00:15:21The name is I don't trust Dario.
00:15:24I don't know if you know this, but you may not know this,
00:15:26but Dario, guess what he said?
00:15:27He thinks that open source models are a danger.
00:15:31He's always pushing for regulation.
00:15:33He does not like open source.
00:15:35He DMCA'd a bunch of people
00:15:37when it came to them accidentally leaking their code
00:15:40and people de-officiating it.
00:15:42He does not like open source
00:15:43and Anthropic doesn't like open source
00:15:45and they want to own everything
00:15:47and they think that AI is too dangerous for anybody
00:15:49but them to run it.
00:15:50And they ultimately just want control over you and your life.
00:15:53They want to leave you helpless and incapable.
00:15:55That is why 90% of all software is going to be written.
00:15:57A hundred percent of all software is going to be written
00:15:59because they don't want anyone
00:16:00to actually be learning and independent.
00:16:02They want everybody to suckle from the teat of Dario.
00:16:05And that is that.
00:16:06And that is what I actually believe.
00:16:08I think that Dario is one of the most dangerous man
00:16:11on all of the AI landscape
00:16:13and that I don't think he has your best interests at heart.
00:16:16(mumbles)
00:16:18I know it's not easy to find developers to work with.
00:16:21Trust me, I hired TJ.
00:16:23So how do you find quality people to work with
00:16:25without spending all of your time on LinkedIn?
00:16:28Well, with G2I, you can go from interview
00:16:29to first PR in as little as seven days.
00:16:32G2I has helped people get placed at meta and one password
00:16:35and they run React Miami.
00:16:37So they're super embedded in the JavaScript community.
00:16:40So if you want to find top talented software developers,
00:16:42check out g2i.co.

Key Takeaway

Anthropic's enforcement of API restrictions against third-party tools like OpenCode is likely driven by unsustainable subsidization costs and a strategic need to lock users into their complete technology stack rather than just their AI models.

Highlights

Anthropic blocked third-party tools like Cursor and OpenCode from using Claude subscription API keys, requiring users to pay significantly more expensive API pricing instead

The restriction was technically always in Claude Code's terms of service since early 2025, but enforcement creates a 'ticking time bomb' that has now exploded

OpenCode is approaching 1 million monthly active users and is considered superior to Claude Code in terms of usability and developer experience

The real motivation may be the massive subsidization costs of AI model development, including training runs, hardware purchases, and employee costs that could be 98% subsidized

Anthropic wants to lock users into their entire stack rather than just their models, as model loyalty alone isn't sticky enough when competitors like GPT-5 exist

The speaker expresses distrust of Anthropic CEO Dario Amodei, citing his opposition to open source and desire for AI regulation and control

Claude Code has significant technical issues including screen flickering bugs that were only '85% fixed' and later rolled back

Timeline

The Controversy Begins

The video introduces the major controversy surrounding Anthropic and Claude Code that has caused significant upset in the developer community over several days. A key error message is shown stating that credentials are only authorized for use with Claude Code and cannot be used for other API requests, which triggered the backlash. The speaker sets up the context by acknowledging that viewers unfamiliar with the situation are living a 'blessed life' before diving into the full story. This opening establishes the heated nature of the debate with references to Twitter users 'screaming' about the changes.

Understanding Anthropic's Subscription Plans

The video explains Anthropic's three subscription tiers: Pro, Pro 5X, and Pro 20X plans. Until recently, users could use their subscription tokens in third-party tools like Cursor and OpenCode to access their allocated or unlimited tokens for code changes. This flexibility suddenly ended when Anthropic enforced that subscriptions can only be used within Claude Code itself. Users who want to continue using third-party tools must now switch to API pricing, which is approximately 10X more expensive than subscription pricing. The change sparked massive community backlash with GitHub repositories filled with users seeking workarounds and expressing their intent to cancel subscriptions.

The Terms of Service Reality

A Reddit post reveals that the restriction was technically always present in Claude Code's terms of service since its release in early 2025. The TOS explicitly stated that subscriptions were only for use with Anthropic's own products, not third-party tools. Third-party tools like Cursor and OpenCode had been using OAuth tokens and forging requests with specific headers and body parameters, creating a fragile integration system. Anthropic had always strongly discouraged this practice and put effort into preventing it, but the lack of enforcement allowed the practice to continue and grow. This created a 'ticking time bomb' situation where many users switched to these tools before Anthropic finally decided to enforce their existing rules.

Anthropic's Official Explanation

The person running Claude Code issued a Twitter statement explaining the enforcement decision. According to the statement, they tightened safeguards against spoofing after accounts were banned for triggering abuse filters from third-party harnesses using Claude subscriptions. Third-party harnesses create unusual traffic patterns without the telemetry that Claude Code provides, making it extremely difficult for Anthropic to debug issues when users have questions about rate limits or account bans. Since these third-party tools lack proper support channels, any problems are perceived as Anthropic's fault rather than the tool's fault. The official reasoning emphasizes that this enforcement is in users' best interests and necessary for Anthropic's ability to provide support and debugging capabilities.

Why Now? The Timing Question

The speaker questions why Anthropic chose to enforce this restriction now, given that tools like OpenCode and Cursor have been using Anthropic's API for months. The timing is particularly puzzling because it comes at the height of 'Claude Code glazing' - maximum positive sentiment and developer enthusiasm. The speaker dismisses the common explanation that it's simply about cost subsidization, where the Pro, 5X, and 20X plans cost Anthropic more than users pay. While acknowledging that some users of higher-tier plans do cost more than their subscription fees, especially the 20X plan, the speaker argues this mathematical explanation is the weakest and most incorrect interpretation. The section sets up a deeper analysis of the true motivations beyond simple cost calculations.

The Model Stickiness Problem

The video references DAX's perspective that models themselves aren't sticky enough as products. While Claude Opus is good, GPT-5.2 is also competitive, with each having different strengths and weaknesses. Anthropic wants to own the entire tooling stack, not just the model, so that when users use Claude, they must use Claude Code exclusively. This becomes especially critical as OpenCode approaches one million monthly active users, growing from a toy project to a major competitor. The growth is attributed to OpenCode's superiority over Claude Code in terms of user experience and functionality. The speaker emphasizes that owning the complete stack provides more value and prevents users from easily switching to competitor models when they perform better for specific tasks.

Claude Code's Technical Problems

The speaker criticizes Claude Code's product quality, pointing out Anthropic's irony as a company that claims software engineering is ending while struggling with their own software. Specific issues include persistent screen flickering that makes the editor look like a 'dance club' and could cause epilepsy, which they claimed to have fixed 85% but had to roll back one day later due to new bugs. Other problems include ANSI escape sequences printing as plain text, the inability to expand code with Control+O while in acceptance menus, and overall poor usability compared to OpenCode. The speaker describes OpenCode as 'absolute beauty' - well-designed by developers who like to code for developers who like to code. The comparison reveals that Anthropic is losing to 'a bald man and a vegan from the Ozarks' despite their resources and claims about the future of software engineering.

The Real Cost Problem (Tinfoil Hat Theory)

The speaker puts on a literal tinfoil hat and presents their theory that the subsidization problem is far more severe than publicly acknowledged. Beyond the subscription tier cost differences, Anthropic faces massive expenses including huge training runs requiring hardware and energy purchases, the short operational lifetime of models as new versions quickly replace old ones (3.5 becoming obsolete when 3.7, 4.5, and GPT-5 variants emerge), extremely expensive employees with outrageous stock grants, and ongoing infrastructure and GPU replacement costs. The speaker suggests that subscription plans might be 98% subsidized rather than just 80%, meaning the true cost of max plans could be 50X more than stated. Additional pressure comes from Nvidia's new Rubin AI GPUs promising 10X reduction in inference costs and 4X reduction in GPUs needed for MoE models, forcing continuous hardware upgrades to stay competitive with ChatGPT, Google, and XAI.

The Strategic Lock-in Motivation

The speaker argues that if Anthropic can create a competitive stack and force users into their highly subsidized plan that only works with their tools, the stack itself becomes the sticky retention point rather than the model. Users lose access to both cheap unlimited tokens and the ability to easily switch to other models when desired. OpenCode is particularly dangerous not just because it allows use of subscription tokens, but because it enables users to easily try GPT-5.2 and discover it might work better for certain tasks, potentially leading them away from Opus entirely. This multi-model flexibility represents an existential threat to Anthropic's business model. The speaker acknowledges being somewhat of a 'fringe conspiracy theorist' but sees genuine trouble, questioning why Anthropic would torpedo the highest amount of developer enthusiasm and constant Twitter praise about software engineering being over unless they were facing serious underlying problems.

Distrust of Dario and Anthropic's Philosophy

The speaker expresses deep distrust of Anthropic CEO Dario Amodei, citing his public positions that open source models are dangerous and his constant push for AI regulation. Dario reportedly dislikes open source and has DMCA'd people who accidentally leaked and deobfuscated Anthropic's code. The speaker believes Anthropic wants to own everything and thinks AI is too dangerous for anyone but themselves to run, ultimately seeking control over users' lives and leaving them 'helpless and incapable.' The rhetoric about 90-100% of software being AI-written is framed not as progress but as a deliberate strategy to prevent people from learning and becoming independent, forcing everyone to 'suckle from the teat of Dario.' The speaker concludes by calling Dario 'one of the most dangerous men on all of the AI landscape' who doesn't have users' best interests at heart, representing a strongly critical perspective on Anthropic's broader motivations and philosophy.

Closing and Sponsor

The video concludes with the speaker inviting viewers to like, comment, and share their perspectives on where the analysis is right or wrong, noting this is unusual for their channel. A brief acknowledgment thanks viewers for helping reach one million subscribers. The final segment includes a sponsor message for G2I, a developer hiring platform that can help companies go from interview to first pull request in seven days, with placement experience at companies like Meta and One Password and deep roots in the JavaScript community through running React Miami. The sponsor integration feels somewhat abrupt after the intense criticism of Anthropic and represents the standard monetization approach for YouTube content creators.

Community Posts

View all posts