Seriously, Anthropic??

I hate to do another video just dunking

on Anthropic, but they keep finding new

ways to just embarrass themselves in the

most public egregious fashions. We've

already seen them doing absurd things

like making sure your build more

depending on what system prompt you have

in clawed code because they didn't want

people using it with tools like OpenClaw

or other harnesses that they don't

support. On one hand, you can kind of

see why they would want that. I've been

running my Open Claw through a separate

service recently and just the pings it

does is about $5 a day of inference

through Opus 47. So I I get that. But on

the other hand, this is absurdity. The

fact that you're using your subscription

in one tool versus another should not

change whether or not they support it.

And even crazier, I shouldn't be build

money when I still have 100% of my usage

available because I used a tool you

don't like. But we can go further here

as Anthropic certainly has. They're no

longer just billing you based on what

tool you're using or what's in your

system prompt. They're now billing you

on what files are in your codebase. Even

that is not as bad as it is. They are

billing you on what file names are in

your git commits. This is a legitimate

report of a user who was paying for a

$200 a month sub for Claude Code that

was using Claude Code and ended up being

build $200 when they hadn't used any of

their usage in Claude Code because they

had a commit that had the term Hermes MD

inside of the commit message. I am just

as tired of talking about anthropic as

you guys are hearing me do it. But when

you [ __ ] up your code this genuinely,

hilariously and pathetically, I have to

make fun of you for it. And that's what

we're going to do now. The level of

incompetence at Anthropic right now,

powered by the disdain they have for

engineers as a whole, is at a level that

is hard to fathom. And as frustrated as

I am crashing out yet again about this,

I think this one perfectly summarizes

the problems I have. But if you want

somewhere better to spend your $200,

maybe somewhere that isn't going to

waste all of your time and money and

burn trust and destroy you, you might

want to check out today's sponsor. I

don't know about you all, but I'm

getting pretty tired of juggling all

these different subscriptions,

especially now that Anthropic banned us

from using my subscription with things

like OpenClaw. What if you want to use

OpenClaw and a VS Code plugin and a CLI

and all these other things that you

might use your inference for? Wouldn't

it be nice if there was one subscription

that covered all of that and good

integrations with all the tools you

already use to use it with? Today

sponsors Kilo Code and they've solved

all of this. Not only does their

subscription work for every different

thing you might want to do, they've also

built their own VS Code extension, Jet

Brains extension, CLI, and more. They

also have my favorite OpenClaw

integration by far. It's full and true

OpenClaw, but without having to spin

everything up yourself. You click the

get your assistant button and they'll

spin up the infrastructure and manage it

all for you. No Mac minis required. As

silly as this quote is, it's true.

They're the only solution that will

cover you from pull request to grocery

list, one subscription will cover

everything from the infrastructure to

the inference itself. Whether you just

want to write some code with the latest

model or try out OpenClaw directly,

there is no better place to start than

soyb.link/ilo.

Let's dive in. I think a little history

is important here before you can really

understand the level of egregiousness

that exists. That a not even existing

file, just an existing name of a file in

your commits is enough to cost you

hundreds of dollars. That's insane. We

need to talk about why Claude charges

for a markdown file that they don't

like. The Max 20x plan on Claude Code is

their highest tier. It is a $200 a month

plan. And this plan is incredibly

generous with the amount of usage you

can get. If you run it in a loop 247,

this $200 a month can get you over

$2,000

of inference. This number is citing the

API prices anthropic charges. If you are

going through their API, traditionally

they charge $5 per mill tokens in and

$25 per mill out on their flagship model

opus 47 as well as 45 and 46. The price

here, this $2,000 is if you look at the

tokens that you utilize as a subscriber,

run it in a loop for the entire month,

you will not hit limits until you've

reached about $2,000 of inference each

week. You're able to do around $500 of

work on the $200 plan. So, if you hit

the weekly limit they provide in Claude

Code, you did $500 of inference. I still

maintain my max 20x plan because I use

it for testing a bunch of different

[ __ ] I barely used it the last few

days. As you can see, you have a session

limit. A session is approximately 5

hours. It starts when you send a

message, which is already iffy because

if it was starting automatically on a

timer, then I could maybe send a message

or two before it resets and then get

more users. But I get why they also have

the weekly limits separately. They have

one for all models. They have a separate

sonnet only one because sonnet used to

be the thing we would use most in cloud

code. Now that opus is the default. This

separate sonnet only section makes

almost no sense at all. Your son uses

counts towards this limit as well as

your weekly and 5 hour session limits.

Like just unnecessarily confusing.

Speaking of which, Claude design is its

own separate thing. Absurd. Genuinely

absurd. But the point I'm trying to make

here is that if you have a weekly limit

and you hit the weekly limit, if you get

100% here, that is approximately $500

of usage. That's a lot of usage for

$200. And if you do that every week for

the month, you'll hit 2,000. That is a

10x discount. As such, they don't want

people to get that close to those

limits. So, if you have a tool that

automatically uses your inference,

they're not going to like that a whole

lot because now they're not going to get

the benefit of people like me who pay

and barely use it. I probably do $50 of

inference a month at most on this plan.

So, they're profiting off of me. But, if

I could use this with OpenClaw, I would

be using it more and then getting much

closer to the $200 I'm paying for. They

don't like that. They don't want that.

So, they're trying to restrict me from

being able to do that. There are some

legitimate arguments here and I will do

my best to steelman them. One of them is

caching. If you're not familiar with how

caching works in AI, I will do my best

to quickly summarize it. You have a

model that has a bunch of parameters.

Parameters are different chunks of text

that point to each other with different

levels of confidence. Like if you have

the capital of the United States is the

most likely thing to come after that is

Washington DC. Or if you just have

capital W Washington, it might point to

a couple different places. is it might

point to DC with high confidence. It

might point to state with slightly lower

confidence and it measures how likely

each of these things is and that's what

the model is. It's a bunch of these

tokens with arrows and vectors pointing

to and from each other to guess what the

most likely next thing is. When you hand

it a history of chat, it has to update

those pointers, those parameters based

on the history of the chat so far. If I

was to say, "What state do you live in?

Is it Washington?" The most likely next

thing is the question mark because

there's indication in the tone of the

other words I used that the most likely