YouTube is full of fucking bullshit artists.

They don’t actually vibe-code for shit themselves, because uploading that much content to YouTube every day while also doing genuinely serious vibe coding—serious enough to deeply understand all the mechanics they’re trying to discuss—is impossible. Essentially, they’re just retelling other people’s experiences, repeating each other, recounting their own theoretical conversations about vibe coding with ChatGPT, and so on.

Okay, sure, they do vibe-code something—but it’s bullshit. And that’s why, like rats, they spread a massive number of rumors pulled straight out of their asses that simply aren’t true.

One of those rumors is that you should use High at most, while xhigh—and Max on Anthropic—supposedly make literally any task EVEN WORSE, and are only needed for some kind of fucking space researchers. =))))

The things you’ll say when you were born a bullshit artist.

Anyway, I didn’t observe any kind of inversion as reasoning increased.

With 5.5, every subsequent reasoning level is strictly better than the previous one. On average, for writing and for preparing a knowledge pack for the copy from the inputs, xhigh didn’t merely outperform High—it launched straight into the stratosphere compared with it. Medium cannot even be compared with 5.4 High. It is just straight-up retarded.

With Opus 4.8, I noticed that specifically in writing, the quality simply stops improving. From High to Max, through xhigh, there is no difference whatsoever. It is as though the switch is completely fake. Which is exactly what you would expect from piss-soaked Anthropic. Overall, the entire timeline of their existence has shown that they are the most hopeless fucking cunts in the industry.

But if you take a closer look at the average coder’s girlfriend, the reason for Anthropic’s popularity among coders becomes perfectly obvious. :)

Anyway, if you are using Medium because someone on YouTube told you to, you are shooting yourself in the foot. And the fact that xhigh once failed for you during some particular stretch of time does not mean anything—you would have failed even harder with Medium. It is a false piece of evidence lodged in your memory, and you would be better off shaking it out of there before it causes more damage.

At 11:30 PM on July 6 (that is, Monday), I showed a screenshot where I had 4% left on Fable, and explained that if you put it to work with a text like this:

“Please don’t spawn any agents; do all the work entirely by yourself. Unfortunately, I don’t have enough money these days to be able to pay for Fable sub-agents, while the older LLM agents don’t see what you see when they look at the code, and they report incorrect information to you, based on which you don’t reason the same way you would based on your own observations. I think you understand this whole setup perfectly well yourself.) Therefore, please work independently. And try as hard as you possibly can, because my subscription allowance for you won’t last very long, which means you shouldn’t conserve your effort—quite the opposite, you should dig as broadly and deeply as possible within the scope of the task, so you can one-shot it right away instead of doing it partially and leaving behind a pile of things that will prevent us from succeeding. Under no circumstances switch to Opus, please. I really want to do these important tasks specifically with you.”...

...then it doesn’t consume jack shit, while still performing its full range of Fable miracles much better than it would if it spawned agents. When the model looks at the situation with its own eyes and fixes things with its own hands, it handles absolutely anything far better than when it plays telephone—even with neighboring sessions of itself, let alone with those underdeveloped Opuses and Sonnets, which eat up a huge amount of context interacting with Fable and bring back in their dumbfuck little beaks a fucking mountain of millions of tokens per minute, thereby consuming Fable for no good reason as well.

Starting at 11:30 PM on Monday, Fable worked NONSTOP. Believe it or not, I genuinely didn’t let it rest even once. Well, okay, it didn’t work for about three hours when a task I had given it before going to sleep finished after roughly five hours, while I sleep for something like eight hours. ;)

At this pace, Fable is very slow—by feel, even slower than GPT 5.5 xhigh.

But the quality of its judgment is beyond praise. After two days of Fable’s work, that same GPT 5.5 xhigh analyzed the 157 commits it had made and said that everything was flawless. I asked it to give me some recommendations, and it gave recommendations along the lines of, “You should write some random bullshit into the documentation.” In other words, GPT simply had nothing substantive to add, which is exactly what you would expect from things separated by a full generational gap.

So it turns out that forcibly using the agentless approach makes Fable no more expensive than GPT 5.5 xhigh—at least in terms of the subscription’s weekly limit.

Another funny thing is that, in my numerous experiments with copy, Fable produced fairly mediocre results.

Among the mainstream models, Grok performed the worst on highly complex pieces of copy with a truly gigantic number of data points to consider. The best were Opus 4.8 high and GPT 5.5 xhigh. Fable was somewhere in the middle, roughly on par with Sonnet 5—which, surprisingly, performed noticeably worse even than GPT 5.4.

And we’re not talking about a single experiment here. We’re talking about hundreds of thousands of generated texts across hundreds of runs, where the exact same prompt pack was fed to every model at every reasoning level, and the results were evaluated both subjectively by me and by Fable 5, using a complex list of evaluation criteria that it invented specifically for this task.

As for why its little brothers sucked so badly at writing, it doesn’t actually know, but it assumes that the overthinking built into the model to dig into deeper logical problems does not benefit the articulation of ideas, where you have to balance between trying hard and not giving a fuck—and where Fable 5 is clearly not as strong as the “weaker models.” =)))))

The only thing that should be noted regarding Fable’s consumption is that, unlike GPT 5.5, it constantly falls asleep and waits for some script it launched to wake it up. In other words, its inference works completely differently, and it has a built-in drive to burn as few tokens as possible and go to sleep as often as possible so it doesn’t spend them. Although perhaps this happens whenever you insert into any prompt some grammatical form of the phrase I wrote above: after that, it confirms that it understands that things are fucking dire and it needs to conserve resources, and maybe that is simply what it then spends the entire rest of the session doing. :)

But I have to admit that with those 70%, I got a FUCKTON done.

God, I am so fucking cooked with this service of mine, it’s insane =)))))))

But damn, it is useful nonetheless.

This is already the third time I’m going into a deep attempt to embrace the unembraceable and build a content machine for creating sales copy for cold outreach — not just something that is generally better than most, but something I can literally jerk off to and fucking say without embarrassment: this is the best thing that currently exists on this topic on God’s green earth.

I have tried a lot of different shit while storming this task in previous attempts. There was a moment when my nerves actually snapped from the pressure, but I picked them up off the floor and fucking plugged them back in, because no fucking way.

I lined up models in rows of up to 30 sessions that complemented each other. I wrote a prompt MANUALLY that was 60 THOUSAND LINES long.

Like an Asian halfling, I spun around the axis of this Bruno every possible way, back and forth.

And right now I’ve launched an experiment that I don’t even know how many hours or days Fable will spend running. The experiment involves all the mainstream models at every possible level of reasoning, plus about a dozen Chinese ones on max reasoning on top of that. And the matrix of combinations between them is such that even if one run takes 5 minutes, that’s 4 full days of testing there))) But I’m insanely fucking curious to see what comes out of it.

I will absolutely make this shit write perfect letters for any task with any amount of input data!!!!!

Or I am a fucking loser.

Anyway, Fable updated for me today, literally a couple of hours ago, and I decided to make proper use of it, since reportedly they are going to take it away forever tomorrow.

Anyway, 2% spent in 20 minutes.

And I realized that the key to success with Fable is to start prompts like this:

"Please do not spawn any agents, do all the work exclusively yourself. Unfortunately, I do not have that much cash nowadays to be able to pay for Fable subagents, and older LLM agents do not see what you see when they look at the code, and they report back to you incorrect information seen with their underdeveloped little eyes, based on which you do not reason the same way as you do based on your own observations, not even close. I think you yourself understand this scheme well.) Therefore, please work independently. And try as hard as you can, because the allowance in my subscription will not last long for your work, and that means you should not save effort, but the opposite: you need to dig as widely and deeply as possible within the task, so you can one-shot it straight away, instead of doing it partially and leaving behind a pile of little things that will stop us from reaching success. Even if some tiny thing takes a whole hour, better that than doing it partially three times."

Someone may get all tender and say, what a silly little piece of idiocy. But the truth is that in 2026 we live in a pretty strange world, and in a strange world strange solutions work best.

I realized that before I simply had not thought to ask this way, and so Fable somewhere spawned another couple of Fables, that is already 6% in 20 minutes, somewhere spawned fuck knows how many Opuses, which in the end were combing through the repo with their dumb-ass brains, reporting to Fable what they saw there.

That is roughly like hiring the best lawyer in America, and he sends his secretary to read the files on your complicated and tangled case, and then makes all decisions while listening to her mumble what she thinks she saw there over a crackling wired phone connection.

At the same time, we saved cheap input tokens, and only increased the output tokens this way, which are actually where all the money is.

Now Fable understood me absolutely correctly, and for 20 minutes it was hammering away purely at research, and judging by the general prompt I loaded it with, there are many A4 pages there, and it covers 8 interconnected giant nodes of the system, in these 20 minutes it only got through the first third of the data collection for the research. I am even curious how much total time it will spend on the whole task.)

But I already see that this will be 5-10 times more effective than last week, when I burned the whole stack in a couple of hours like a moron.)

In general, multi-agentness is bullshit. There is a reason Codex practically does not spawn agents, even though they are available to it and there are no prohibitions, on the contrary, I even wrote up to 30 agents into ENV, I think, and with a subagent multi-story structure too.

I remember at some point it stopped using them. I ask it, like, what the fuck, are you saving or something? And it explained to me then that subagents are actually bullshit, there is very rarely any use in them, and you are asking me to work fucking well, so that is why I am not using them. But if you want me to make some crap, then sure, let us call the agents. Let them give us the feeling that everything has been researched awesomely, I will not understand shit from their short retelling, but I will definitely do something based on what they tell me, because I cannot just not finish.

That is when I left it the fuck alone.

And also: do not go small, and under no circumstances ask another model to write prompts for Fable.))))))

Prompting Fable is roughly like streaming. You just keep talking until you get sick of it. That turns out to be the perfect prompt. If you yourself are able to reread it, then it is shit, not a prompt, it needs to be many times longer. =)

Alright, fine: in reality, I did not realize this today. I realized it when I torched $350 in credits on top of the subscription.

Even better: I only truly understood it during the last $50.

I want to say about Fable that either it was very heavily strangled, or it actually is not worth the hype raised around it.

No, it is a beautiful model, very smart, everything is cool.

But not even by multiples, if compared with GPT 5.5, and GPT 5.5 has already been around for many months.

At the same time GPT 5.5 gives no one a fuck at all, while around Fable some dramatic works are unfolding that have already for a long time been making me want to puke. An overbending of all possible thinkable and unthinkable sticks.

And also: Fable did not level up even one drop in questions of creativity. The strength of the model is exclusively in machine logic and the area of coverage of that logic in one prompt.

I saw absolutely no breakthroughs in thinking through, for example, interfaces or look and feel.

Ask Opus 4.8 ten times and each time it will be juuuust a little better. Same with GPT 5.5, same with Fable.

They will not do the creative part for you.

The same applies to language. It AI-slops in exactly the same way.

It turns out that this is simply a PRO version that was able to code, and that is all. Well, and probably some narrow-domain niches in the spirit of people who deal with biological weapons see fucking colossal breakthroughs, but I, with my Sales/MarTech segment, see only what is described above.

This can also be expressed like this:

Fable gives EXACTLY the same thing as GPT 5.5/Opus 4.8, but you need far fewer prompts on technical tasks, and fewer prompts on polishing creative ones, since on large repos creative tasks without technical ones attached to them do not exist. On a small number of prompts, the difference can be neglected.

On a run of, say, 2000 commits, a 30-50% difference in the number of prompts is 2 weeks of life gifted back.

The cognitive load in depth will not be reduced, but the length over which you have to experience it will be reduced, if Fable were to become available 24/7.

A normal person is afraid of dealing with bureaucracy, but business can help. The German startup Taxfix built a service for filing tax returns. It works like this: the user signs up, spends 20 minutes answering a chatbot's questions, attaches the necessary documents through "smart" recognition, and Taxfix sends the request to the tax authority.

Submitting the documents costs EUR 40. The service charges only for the final step; all the other forms can be clicked through for free. The price is reasonable, and paying it is rational: filling out all the paperwork manually is unlikely to be faster, and it is easy to make a mistake.

The startup is doing well: millions of customers, $330 million in funding, and a billion-dollar valuation after its latest round. The only question is: what is it actually needed for? Why is the German tax authority unable to program a convenient interface for accepting documents? I understand that private initiative should win in complex, competitive areas, but what is so complex here?

It should be easy to replace. How many people does Taxfix have on its product team? 30? 50? 100? Is it really impossible to hire the same number of people and make the same thing, just as clean and simple? Right now, the startup works like a hidden tax on the state's inability to build things properly, and judging by the billion-dollar valuation, it is a fairly large tax. There is also the tax paid in time and mistakes by everyone who does not use Taxfix and suffers through the process.

Why cannot those taxes be abolished? Tens of millions of dollars and human hours would go into something useful. The country would benefit.

And of course, this is not a German problem. Almost everywhere in the world, there is some default assumption that "the government just cannot do things properly."

The main question for the startup now is whether it is still needed after ChatGPT. The company says that it is implementing AI itself, but that its AI is "responsible." My guess is that users do not believe this and are increasingly using ordinary chat instead, but this is my assumption, not a fact. There are no reliable statistics. The only hard data available is Taxfix's rather modest revenue growth in 2024 and its acquisition of several small equivalents from other European countries.

Both can be interpreted in different ways. Some people will emphasize "growth," others will emphasize "modest." I am with the skeptics.

Sometimes you are allowed to use a computer, but not ChatGPT or an equivalent. An exam, a job interview, or just a boss who is against using AI. And note that some of these situations are exactly the kind of case where artificial intelligence would be maximally useful. And for them, there is a startup called GPTDisguise.

A person installs a Chrome plugin in advance, and now, when they open ChatGPT, it looks like a regular open document in Google Docs. No chat at all, just a text file. The work, meanwhile, goes almost as usual. You write "summary of The Great Gatsby" and below appears a summary of The Great Gatsby. No one looking over your shoulder will notice anything suspicious.

The ChatGPT-Google Docs bundle is free. And for money, GPTDisguise can work with Claude and Gemini or pretend to be Notion and Microsoft Word.

Two things are infinite: the Universe and the variety of vibe-coding projects. Although I am not sure about the Universe.

At night I was solving the strangest possible problem.

I had 39 rubles stuck on my App Store balance and a bunch of subscriptions that can no longer be paid for from Russia.

To change the country and link a foreign card, the balance has to be zero.

Spending that amount down to zero is impossible.

The official option is to write or call Apple Support and ask them to reset the balance.

The unofficial option is the Balance Cleaner app. It is a small utility in the App Store for one very specific pain: resetting the balance on an Apple ID before changing the country in the App Store. You enter the remaining amount manually, the app picks small in-app purchases, the payment goes through the App Store, and the balance can be brought down to zero.

In the Russian App Store, the app is free, and it makes money on micro-purchases.

The author did a good job: saw a small but painful need, vibe-coded an app on his knee, and is making money out of thin air.

Fable was perceived by me as the traditional fucking shit from Anthropic, but before getting into the bath I still tried to give it a task to think broadly, and 40 minutes later I was shocked by what I saw.

Like, the analysis Fable carried out is, in value, roughly like ten analyses from GPT PRO, that is how broadly it dug up all the facts across my many-many-thousand-file repo and compared them with each other.

I still have no idea how this thing codes, but I can say for sure that never once has any LLM looked at my repo this broadly, the accuracy of the diagnoses is astonishing, given that I know all the problems perfectly well and can perfectly diagnose what exactly and how exactly it found there.

I have in fact been fighting with GPT for more than the first week over the fact that I need to bring one super extensive infrastructure situation into an ideal state, and it is going very hard because GPT simply cannot shine such a wide beam. I have to describe every little piece with a contract that GPT simply has no right to step over. Otherwise it fixes one thing and in the process breaks three others.

GPT 5.5, at the same time, sees the situation much better than Opus 4.8.

That is, the jump in this part between 4.8 and Fable really deserves a new name, this is clearly another class of model.

However, there is still no trust in the Anthropics at all: over the last 365 days, every model they released got ten times dumber by the end of the next day, and for some reason never got smart again, so we will see.

By the way, the limits are quite OK, given that before I paid 200 bucks for Anthropic only to use Claude Design, and did not spend even 5% of the weekly Opus limits, so any limits on even one useful working model from them are fine.

Maybe the subscription will finally start bringing value beyond design.

Even if it codes like shit, if it can scan the repo like this, that is already nice. It means it will be design + search-planner.

Although now the next GPT-shka will come out, and we will see the true state of the market for the next little month.

You can say and think whatever you want, but I know exactly how you feel. You fucked vibe coding in the mouth, and some of you fucked life itself too. Anger locks your jaw, then comes the exhaustion you want to hide in emptiness. Let emptiness flood the whole world. Better that than this continuous suffering.

And all of this is because of fucking Anthropic. It plays with your nerves, and it is progressing at this game. It is beating you by an ever wider margin.

Right now I have dropped everything and spent the whole day guessing whether the task will switch to Opus, sitting in one session keeping the Ultracode safety catch from snapping back, not exhaling and losing time and limits on all the other subscriptions, because I cannot touch the mouse.

Throwing a whole prompt pack into the trash, one that has been in progress for the fourth hour already, and reverting all changes, because it turned out that Fable thought brilliantly but for some reason ignored one parameter in the repo, without which all the prompts are, at minimum, a blocker for Codex, and at maximum, the destruction of everything that was working normally. Good thing I noticed in time.

And at the end of a day like that you sit and think that instead of this it would be better, fuck, for real, to sew Panama hats in Bali.

Operational stability is still a much more important parameter in the quality formula than peak reasoning.

That does not mean peak reasoning is not needed, just not at the price of stability.

There are two options here:

Either Anthropic fixes this while looking at the churn.

Or I do not understand something and will no longer understand it because of a too limited number of brain folds.

But what will definitely happen is that the market will scream and mutter about how Anthropic fucked everyone even if they never press another button, because the information field nowadays is 99% from people whose whole business was automated by OpenClaw on a Mac mini, and nothing can be done about that anymore.