Why I Cancelled My Claude Code Subscription

0:00I was a Claude Max subscriber, I paid $200

0:02a month, I loved it.

0:03I was a beast power user, using hundreds

0:06of millions of tokens per day sometimes.

0:08I can easily say that

0:09just over the last 12 months,

0:11Claude Code went from pretty neat to "OMG

0:14can't live without it".

0:15Something as a builder and solopreneur I

0:17genuinely couldn't do without.

0:18And that is exactly why I

0:20canceled my subscription.

0:21Because I realized there was something

0:23seriously wrong with a

0:24dependency on Enthropic.

0:25Let me explain.

0:26You see over the last 30

0:27days of my subscription,

0:29on two of them I surpassed 700 million

0:31tokens per day, and other

0:32days were well on their way.

0:34If we do all the math here

0:35with the general API cost,

0:37those two days alone

0:38would have cost over $50,000,

0:40around $2,000 if you

0:42factor in all the token caching.

0:43But we'll get to that later.

0:44But even with the discounted cash tokens,

0:46we're still looking at

0:48approximately $1,800 per month,

0:49just for two days of usage, and my

0:51subscription only costs $200.

0:53Something is seriously

0:54wrong with the economics here,

0:55and I don't want to be hooked on Cloud

0:56when the other shoe drops.

0:58How was I getting all

0:59of this for $200 a month?

1:00Someone asked to foot

1:01the energy bill here.

1:02AI is not free.

1:04Yet the more I use Claude Code, the more

1:06my work and business depended on it.

1:08Until I finally started thinking about

1:09what would happen if that $200 plan

1:12disappeared next week.

1:13What if they changed the

1:14limits or the product itself?

1:15At that point, how easy

1:17would it be for me to leave?

1:18And over time that

1:19question started bothering me.

1:20A lot.

1:21Because a large part of my life was

1:22revolving around Claude.

1:23Fortunately, there's a way out of this.

1:25Today we have open-weight models that are

1:27not only surprisingly good,

1:28but they offer results

1:29comparable to Opus and Fable.

1:31Meanwhile, there are more and more

1:33open-source agent frameworks and

1:34harnesses that are model-agnostic, which

1:36allow you to get

1:36creative past the handful

1:38of models offered by Claude, or OpenAI, or

1:40any American frontier family of models.

1:43So I decided to try something.

1:44I want to see whether I can replace my

1:46$200 Claude subscription

1:47with open models and a

1:48smarter workflow, without giving up the

1:50things that made Claude so damn useful to

1:53me in the first place.

1:54Let's start with the token inference.

1:56As I said, those two days alone would

1:57have cost approximately $1,800.

2:00But that's only two days of my 30-day

2:02subscription that I pay, in total, $200.

2:05So I started digging into

2:06that because it seemed strange.

2:08The API prices are public.

2:10If you took the amount of input and

2:11output tokens from those days

2:13and priced them as normal API usage,

2:15well, you see these high

2:16numbers in the tens of thousands.

2:17So either their API margins are obscene,

2:20or they're burning VC money

2:21to subsidize user inference,

2:23get us hooked, and hiding the

2:25actual cost of delivery here.

2:26It's probably a little bit of both.

2:28They have a good product.

2:29People want to pay for it.

2:30Claude Code has become one

2:31of the most useful products

2:32some of us have ever

2:33paid for in our lives.

2:34We use it constantly.

2:36Someone like me, I could sit down on a

2:37weekend, start building something,

2:39and not stop for 20 hours straight.

2:42The more you use it, the more you got

2:43from that easy $200.

2:45That's an incredible

2:46deal for a heavy user.

2:48But there is a flip side.

2:50Once you get used to working that way,

2:51you build your entire workflow around it

2:53and you become dependent.

2:54You stop thinking about

2:56the cost of every request.

2:57You stop worrying about

2:58which model you're using.

2:59You just open Claude

3:00Code and keep working.

3:01And I had definitely

3:02gotten to that point.

3:03Lazy with my AI usage because the Claude

3:06Code harness itself,

3:07with the subsidized

3:08inference, allowed me to just work

3:09and not have to really think about

3:11anything under the hood.

3:12So when I started looking at how much of

3:13my actual work was being done with Claude,

3:15I started wondering what

3:16happens if the deal changes.

3:18Because we've seen AI products or

3:20anything in tech

3:21change their limits before.

3:22A plan that looks generous

3:24today can look very different

3:25once the economic reality of

3:28running a business kicks in,

3:30which is often what happens when a

3:31company is about to IPO.

3:33But that's another story.

3:34So I'd already seen how

3:35quickly my usage could explode

3:37with over 700 million

3:38tokens in a single day.

3:39That's just what my

3:40workflow started to look like.

3:41So I had to ask myself

3:43the uncomfortable question.

3:44If Anthropic changed the price tomorrow,

3:46or tighten the limits, or

3:47change the way Claude Code worked,

3:49how much of my workflow would I have to

3:51rebuild or start from scratch?

3:52And that question bothered

3:54me a lot more than the $200.

3:57So I used Claude Code to do a lot.

3:58Build landing pages, sales pages,

4:00automated meta ad campaigns,

4:02generate visual assets for videos or

4:04presentations with

4:05emotion and hyperframes,

4:07and handle parts of my content workflow,

4:09including repurposing, metadata

4:10generation, et cetera.

4:11I did a lot.

4:12And then there were my side projects.

4:14I had a project I was

4:14working on with my son,

4:15where he created a world, built a story,

4:18and we turned it into a

4:18book, and a YouTube channel,

4:20and an educational alerts or

4:21read game on its own website.

4:23We made our own vibe-coded Patreon.

4:25And Claude was sitting

4:26right in the middle of that.

4:27If I didn't have the subsidized free

4:29inference, essentially,

4:30to build that, I wouldn't.

4:32Because I wasn't

4:32guaranteed to make any money.

4:33It was just a fun project.

4:35And it was with this project with my son,

4:36I realized I would open Claude Code before

4:39I'd even think about

4:40whether there was another

4:41way to get the job done.

4:42I already knew how to use it.

4:43The tools were there.

4:44Context and memory was there.

4:46I built up all my workflows around it.

4:48So starting something new was easy.

4:50And that convenience

4:51adds up as an ignorance debt

4:53that eventually we must pay back.

4:55Because you've stopped

4:56questioning the tool,

4:57because it just works.

4:59And after enough time, you

5:00inevitably feel locked in,

5:01whether you realize it or not.

5:03The cost of switching

5:04just becomes too high.

5:05And that's what Anthropic wants.

5:07That's what any business wants.

5:08It's good business.

5:09Because it turned into even the thought

5:11of me canceling my Cloud subscription.

5:13It was more than just

5:14cutting another subscription

5:15from my life.

5:15I'm not just going to

5:16stop building with AI.

5:18I have to figure out an alternative.

5:19There's a cost to switching.

5:21Some of my workflows

5:22would be easy to move,

5:22but some of it would not.

5:23And truthfully, I didn't

5:24like how little I'd actually

5:26thought about that before.

5:27I'd spent all this time

5:28making Cloud more useful to me,

5:30while quietly making

5:31my own workflow harder

5:32to separate from it.

5:33And that's a weird position to end up in.

5:35You start with a tool

5:36because it's convenient,

5:37then you build around it.

5:38And eventually, that

5:39convenience becomes the reason

5:41you don't want to leave.

5:42It's like the relationship

5:43you know you need to end,

5:44but it's just too difficult.

5:46So eventually, two things

5:47were really bothering me--

5:48the future economics of the subscription

5:50and the company, the vendor lock-in,

5:52the fact that I built

5:53so much around a system

5:54that I didn't really control.

5:56So at this point, I asked myself,

5:58I don't want to just be stuck with Cloud,

6:00and I don't want to just resign to that,

6:02especially with all

6:03the geopolitical drama

6:04Anthropic has been cooking up.

6:05So I started looking for a way out.

6:07Now, the question arises, what the hell

6:09do we use instead of Cloud or OpenAI?

6:11Because I'm talking about Cloud here,

6:13but the truth is this

6:14applies to any AI provider,

6:16especially one that

6:17doesn't respect your privacy

6:18and is giving you something for free.

6:20Remember what they say,

6:21there is no such thing

6:22as a free lunch.

6:23And if lunches were tokens, well, I've

6:26gotten a lot of free

6:28lunches with Cloud code.

6:29But the answer to this

6:30question is not to replace Cloud

6:32or Codex with one thing.

6:34And that's how I was

6:35thinking about it the wrong way

6:36at the beginning.

6:37I was looking for a single tool that

6:38could replace Cloud.

6:39Oh, is Codex better?

6:40Is Gemini CLI better?

6:42But that doesn't really make sense,

6:44because why should one

6:45model or one family models

6:47have to provide everything?

6:48If I'm writing code,

6:50maybe I want just one model.

6:51If I'm planning a

6:52project, I might want another.

6:54And this is where things got interesting,

6:56because the options have

6:57changed a lot since I first

6:59started using Cloud code.

7:01Open-weight models now

7:02can do some serious work,

7:03from Kimi K3 to GLM 5.2, GLM 5.3,

7:08DeepSeek V4, Q1 3.8,

7:10and you can connect all these to open

7:12source agent harnesses.

7:14And a tool like Open

7:15code, which has quickly

7:17become my Cloud code replacement,

7:19is what they call model agnostic.

7:20I can connect any

7:21provider, any model into there.

7:22And it's completely

7:23compatible with my Cloud code agentic

7:25workflows.

7:26So once I started

7:27thinking about it that way,

7:28the whole setup changed.

7:29I was able to start using

7:30Kimi K3 as more of a planner,

7:33an orchestrator.

7:34It's more expensive, but

7:35it's super intelligent.

7:37And it's actually a

7:37lot friendlier and easier

7:39to work with than Fable.

7:40Meanwhile, GLM 5.2 is cheaper, could

7:42handle a lot of the coding.

7:43And it's a very good

7:44creative writer as well.

7:46And for even simpler jobs, DeepSeek V4

7:48Flash may as well be free.

7:51It's fast, intelligent, and while I

7:53wouldn't want it planning my projects,

7:55it can handle the manual tasks perfectly.

7:58And getting comfortable with this

8:00meant that I could

8:01move away from the belief

8:02that I need to Cloud code

8:04to achieve the productivity

8:06that I desired.

8:07But there's another

8:07advantage I want to double click on.

8:09And that is actually

8:10choosing where the models come from.

8:11Now, as I was talking about, when

8:13I'm creating something

8:13for my business content,

8:14like what does the privacy really matter?

8:16Well, in a lot of ways,

8:17the output itself, whatever.

8:19Privacy, no privacy, whatever.

8:20But there's the actual creative process

8:23that comes from me, my

8:24unique way of thinking

8:25as a human being that is

8:27being profiled by tools

8:28like Cloud Code, Codex, Gemini.

8:30They're learning my psychology

8:32and making a profile about me.

8:34And once you start going deep into that,

8:35it's kind of freaky.

8:36So there's another reason to shift

8:38from these proprietary frontier services.

8:40And that begins to align with the

8:42inevitable next step,

8:43if you're a builder with AI, is what

8:45you can start using these harnesses

8:47for inside your own house.

8:48I started messing around

8:49with home security systems,

8:51with Raspberry Pis in the

8:52living room hosting video game

8:54emulators.

8:55And once I hit the stage, I realized

8:56I don't want an

8:56anthropic in my home network.

8:58No, thank you.

8:59And the models I can host

9:00locally just aren't there yet.

9:01So that's why I use

9:02Venice as a provider of all

9:04of my favorite open weight models.

9:06Because they have zero

9:07data retention agreements

9:08with the providers.

9:09Meaning nothing is

9:10being stored by the provider

9:12of my inference.

9:14No profile is being built.

9:16And even if there's a

9:16nefarious entity inside the GPUs

9:19that my provider doesn't even know about,

9:21they're not going to

9:21know it's coming from me.

9:22So this started

9:23creating a comprehensive view

9:25of how I want to

9:26approach my AI setups now.

9:28Instead of opening

9:29Cloud Code and asking it

9:30to do every little thing,

9:31I can now build workflows

9:33that use different

9:33models, different tools,

9:34different providers for

9:35different parts of the job.

9:37And that brings us into a

9:38topic that most Cloud Code users

9:39have probably not thought about much.

9:41And that is smart model routing.

9:42I had no choice but to

9:43start digging into this.

9:44I'd gotten into the

9:45habit of treating Cloud

9:46like the default answer to every problem.

9:48So once I stopped doing that, the

9:50economics actually started

9:51to look very different.

9:52I could have one model think

9:53through what actually needs

9:54to happen and then hand

9:55off the plan to another model

9:57for implementation, a cheaper model.

10:00And if it's not a coding thing,

10:01I could use even simpler models.

10:02And then when I need a

10:03serious final review or audit,

10:05then I bring the big boys back in.

10:07So the most interesting part

10:09of something like smart model

10:10routing is that it can be

10:11handled in the workspace itself.

10:13So what might not be

10:14obvious as a Cloud Code user

10:15is that your AI agent can

10:17make calls to other models

10:19with instructions, with a

10:20system prompt, whatever.

10:22And what's really interesting about this,

10:23if you've only used Cloud Code or Codex

10:25and you've never

10:25built your own workspaces,

10:27is that the workspace

10:28itself can handle a lot

10:30of this smart model routing.

10:32The better your

10:32instructions are in AgentsMD,

10:35the smarter your agent can

10:36be in sending certain tasks

10:38to certain models.

10:39This will not only give you better

10:41performance and results,

10:43but it will save

10:43money, which is important

10:45when we're no longer

10:46using the subsidized inference

10:48by Anthropic.

10:49A well-structured

10:50workspace means my agent is going

10:52to be smart about the

10:53tasks it does versus the tasks

10:56it sends subagents to do.

10:58So as an example, I could

10:59use Cloud Code for a task.

11:01I could use Open Code for a task,

11:02and they might use a

11:03completely different amount of tokens,

11:05even if I'm using the same model.

11:07So let me explain.

11:08If I throw everything

11:09at Cloud Code and say,

11:10hey, figure this thing out, just do it.

11:12It'll probably use a lot of

11:13tokens in a fresh workspace

11:15with no instructions.

11:15It'll make its own instructions.

11:17It's so smart, it'll figure it out.

11:18It'll probably do a good job.

11:19But it's gonna burn

11:20through way more tokens

11:22than if you use the

11:23same model, Opus, Fable,

11:25in a tight workspace with

11:27instructions, with skills,

11:28with subagent profiles, et cetera,

11:30because it has context to work with.

11:32Not only that, but

11:32within the model workspace,

11:33you can integrate model routing.

11:35So if something needs to

11:36be done by a different agent

11:38in your instructions, you simply say,

11:39when we do this thing,

11:40send it to that model

11:41from that provider,

11:42here's the API key, done.

11:44You could do this with Cloud Code,

11:45but why would you even think to do that

11:47when Cloud Code is just so good

11:48and you're not paying

11:49for the actual inference?

11:51And a whole generation of AI users are

11:53gonna be handicapped

11:54when reality comes knocking.

11:55And I don't wish that for you,

11:57so I'm glad you're watching this.

11:58And now is probably a good time to

12:00subscribe to my channel

12:02if you do wanna learn

12:03how to make workspaces

12:04that are not dependent on Cloud Code

12:05because that's what I explore in my

12:07videos on the channel.

12:07So thanks for hitting subscribe.

12:09So once I started

12:09understanding the different economics

12:11of running different

12:13models in a workspace,

12:15I wanted to know how far I could take it.

12:17Could I actually

12:17rebuild enough of my workflows

12:19that I wouldn't miss Cloud?

12:21The answer was way more obvious than I

12:22could have expected.

12:23So I decided to actually test it out.

12:25I've canceled my Cloud subscription.

12:27It's time to move on.

12:28I'm running a Venice Max account,

12:30which is also $200 a month,

12:32and it gives me $225 worth

12:34of credits with their service.

12:35But there's over 300 different models

12:37that I can choose from

12:38from the top of the

12:39line open weight models

12:40like human K3 to GLM 5.2 to

12:43DeepSeek V4 Flash, et cetera.

12:45There's a lot to play with here

12:46that will allow me to save

12:48money and retain performance.

12:51It just means

12:51sacrificing the convenience.

12:53My goal is not to

12:54recreate Cloud perfectly,

12:55but it is to get similar results

12:58without having to spend a lot more money.

13:00The whole point of my

13:01experiment is to find out

13:03where the gaps in my own

13:04workspaces and workflows are.

13:06I have to be present

13:07with every single word

13:08and every single

13:08prompt because all of those

13:10are costing me tokens, which cost money.

13:12If a task doesn't

13:13need the strongest model,

13:14I shouldn't send the task there.

13:15If a bunch of context is

13:17gonna be used, can it be cached?

13:18If I'm recycling

13:19instructions, I need to make skills

13:21and put them in my

13:22behavioral context files.

13:23This is the stuff I'm most interested in

13:26because this is the true

13:27architecture of owning your own AI.

13:29Sure, I might not be

13:30running the actual intelligence

13:32on a machine I own, but theoretically,

13:34in the not so distant future, I could.

13:36And more importantly, as

13:37long as there are vendors

13:38out there willing to provide

13:40for me the models at a cost

13:41that I wanna use, then the

13:43most important aspect for me

13:44is owning the workspaces themselves

13:46because they can be mine

13:48and no one can take those away

13:49from me, no one can shut those off.

13:51I'll just find another model provider.

13:52We live in a world of markets.

13:54Find the service

13:55provider that works for you.

13:56This way, if a particular

13:58tool or model or provider

13:59stops working for me,

14:00the rest of my workflow,

14:01the rest of my life with AI

14:03doesn't have to come down with it.

14:05I don't know yet how far this will go,

14:07but that's what the

14:07next few weeks of my life

14:08are gonna all be

14:09about, not using cloud code

14:11and moving into agentic

14:12workspaces that I control

14:13with inference that I'm paying for,

14:15every single prompt accounted for.

14:18A responsibility that I

14:19have chosen to stop outsourcing

14:24can open weight

14:25models, open source agents

14:26and smart model routing

14:27replace what I was getting

14:28with cloud code, or am I going to end up

14:31resubscribing to

14:31Claude tail between my legs

14:33in a few weeks from now?

14:34That's the experiment.

14:36Stay tuned because I

14:37will be posting updates,

14:39letting you know how it goes.

14:40Finally, if you're interested in learning

14:42how to build these

14:43agentic workspaces yourself,

14:45I invite you to check out

14:46the AI Captain's Academy.

14:47There is a whole

14:48curriculum inside for understanding

14:49from the ground up how AI agents work,

14:51how to select between

14:52models and providers

14:54and agent frameworks, et cetera,

14:55how to approach building your agents,

14:57including knowledge, infrastructure

14:59and context engineering.

15:00We do two calls a week.

15:01One is a workshop, one is a

15:02questions and answers session.

15:04Come say hi, being a

15:05community of builders

15:06that use AI that love AI like you and me,

15:09and we share our

15:09solutions to these problems

15:11and challenges we face,

15:12such as replacing cloud code.

15:14So if that sounds interesting to you,

15:15check out the link below

15:16and I hope to see you there.

15:17Thanks for watching and I

15:18hope this has inspired you

15:19to reconsider your own

15:21Claude or OpenAI subscription.

25:00[Music]

27:00[Music]