> Numbers like that buy a model a real migration effort.
Such a silly choice of words. I wish the human directing the LLM writing the article put some effort into rewriting the worst examples of LLM style.
> But it did extremely well, and the promise was immediate and specific: builds finishing in less than half the wall-clock time, at 27% lower cost, scoring at or above our incumbent on completed work.
The way the LLMs write (Claude perhaps?) With short phrases separated by colons, commas or full stops, is so poor and frustrating.
There some good insights behind this article, so it's worth reading, for example below, but it isn't easy to read.
> Earlier GPT models cached implicitly on partial prefix matches, which gave decent hit rates for free. GPT-5.6 dropped partial-prefix matching:
Can we get over the detective work about if the text was written by LLM or not in 2026 already ? This is a lost cause, and we could instead focus on substance over syntax.
Not OP but my frustrations come from it being impossible to ignore and outright distracting.
I've found the same thing showing with Claude-coded/designed front ends that overuse the same semi-monospaced fonts, Blue/Yellow/Red palette and rounded corner borders. It isn't that it is bad, but it often isn't fit for purpose.
You're right it wont change anything, but authors shouldn't be surprised when people who care about their time/attention comment on low/no effort pieces.
No but it's a useful shorthand to describe a type of bad writing.
I also think that people should focus on substance and not if AI was used, but AI writes like shit and I find myself retching a bit when I have to read long AI-written documents. Do they say something useful? Maybe, but when my eyes are glazing over because it's just so exhausting trying to parse what's written, I can't tell.
I certainly think less of people when they have such poor taste that they think writing like that is acceptable.
Slop is slop. When the “it works but isn’t great” phrases end up slipping into a strong conceptual core, it compromises the perception of the ideas.
Perhaps our AI will cater to us by rewriting the content we read, and each of us intermediate all communication with systems that make that slop bearable.
Or perhaps, we learn that we kinda still need to give a shit when writing to land on the perception we’re trying to create within our readers
To me it's a useful signal not to read an article that someone didn't bother to write.
Which is a shame as real insights are buried inside some of these articles, which if the author bothered to write in his own words could have reached an audience that would have appreciated them.
Writing is one of the areas where I want no LLM involvement.
More often it's the difference between finishing something and not finishing it - so often LLM's are helping them reach an audience that would appreciate them, even if that audience doesn't include you
The problem is that the second you suspect something is written by AI, its a pretty good signal that 50-80% of the text is empty of meaning. Maybe that will change, but LLMs are terrible and inefficient writers.
Only so much time in the day, its a quick signal to not waste anymore of it.
Correct. AI == Credibility hit and it's increasing as more humans get used to feeling they are AI slop consumers, not worth the time for genuine human engagement. Human engagement costs are increasing. Amazing to read/watch.
It's not about figuring out if it's LLM written though. The style is hard to read and annoying. With the kind of sentences GP was talking about it's actually harder to get the substance.
Yes, as soon as models come out that can write properly, we'll all instantly get over it. Until then we'll be having this discussion over and over, as many times as it is necessary.
No. As a human I like reading human written text over computer written text. I want something a human composed with thought put into it. Not something a human tried to save time with by having the machine write it.
Evaluating substance takes time - perhaps more than was invested in the article to begin with. So these tells are very distracting because as soon as I see them I wonder if the person who prompted the LLM even bothered to read the output. If they haven't, then I certainly shouldn't invest the time to determine if there is any substance.
I have a counter proposition: don't fall for this constant suggestion that LLMs are an unavoidable future would you leave the techbros alone now pretty please, relentlessly keep reminding that we still don't think it's acceptable so people don't start to think this is okay since nobody complains anymore.
I appreciate these comments, they save me time for procrastinating elsewhere.
I agree with this sentiment: it’s not inevitable if we relentlessly ostracize obviously LLM posts
And let’s be real: I had a post this year that was #1 on HN for a while, and an LLM “wrote” the whole thing, but it was very much my writing style and NO ONE called out the post as LLM slop. If you use an LLM correctly for writing, it’s not detectable. It seems that most folks don’t go through that effort.
> The way the LLMs write (Claude perhaps?) With short phrases separated by colons, commas or full stops, is so poor and frustrating.
Yup llmish (from now on it's called "llmish") sucks.
But I'd say: at this point it's probably trivial to write a browser extension that detects llmish and that rewrites the worst sentences: from llmish to something less irritating to read. Heck, I could spent tokens on that: an extension that changes on the fly llmish found on webpages.
Also I'd say there's typically no swearing at all in llmish: llmish is too politically correct for swearing. So the rewrite could maybe also use a few "offending" words.
Offending words that, btw, are not going to go well with Gen Zers. Poor Gen Z... They've been raised with the state and its institutions (like school and then universities) hammering them with the notion that they were precious little unique snowflakes and now they arrive on the job market only to be told they've been pre-emptively replaced by AIs. And because they cannot stand a single curse word (because it's "offensive to minorities" or something), they'll be driven off by text rewritten to contain curse words. So they're condemned to read the bland, dumb, AI-generated llmish for the rest of their lives.
We run a lot of varied, tiny, simple workflows that were previously running on 5.4-nano and mini. We transitioned them to 5.6 and noticed exactly this range of improvement across the board. In a few cases, we had improvements in classification.
I think a lot of people miss that for many companies, a model upgrade like this is basically a one liner.
Even if you have an amazing model router architecture (which we do for our golden flows), it’s just not worth it. Not to mention reliability and so on
> Ploy’s agent builds and edits real marketing websites. It plans a page, reads the codebase, writes components, generates imagery, screenshots its own work, and decides when it’s done. That job description sets a very high bar for a model, and we test every frontier release against it. For the four months Opus held the default slot (first Opus 4.7, then 4.8), nothing we tested beat it.
Well, unlike OP I haven't run a rigorous test, but I still would expect Fable to be significantly better at building marketing websites than Opus. It sure is way better at building decks.
The cost reduction is impressive, but I think consistency matters even more for production agents. I'd be interested to know whether prompt engineering or tool-calling workflows had to change significantly.
The problem I always run into with subagents is that they are isolated. This is a double-edged sword, as it keeps context down and lets them "focus", but it often means they must do their own research to continue to do work given to them, which eats uncached tokens.
So depending on how heavily agents are used on what tasks, it's entirely possible that you get worse work for more cost.
This is a feature if your goal is to obtain many samples. Independence is critical. This makes it easier to accurately model the uncertainty of a decision.
I feel Claude Code has added (and removed?) a feature that forks a subagent from the parent context, so it’s still isolated but it’s more of a continuation of what you were doing in one narrow direction and then it dies. Rather than a blank slate with a prompt of what to do.
But what users prefer? Given this is for marketing, which results produce more conversions? From the examples shown, personally I strongly preferred Claude Opus in all cases.
Thank you for a dense informative article with practical takeaways. This was an easy read and it reinforced the importance of some concepts in LLM based pipeline design.
> Numbers like that buy a model a real migration effort.
Such a silly choice of words. I wish the human directing the LLM writing the article put some effort into rewriting the worst examples of LLM style.
> But it did extremely well, and the promise was immediate and specific: builds finishing in less than half the wall-clock time, at 27% lower cost, scoring at or above our incumbent on completed work.
The way the LLMs write (Claude perhaps?) With short phrases separated by colons, commas or full stops, is so poor and frustrating.
There some good insights behind this article, so it's worth reading, for example below, but it isn't easy to read.
> Earlier GPT models cached implicitly on partial prefix matches, which gave decent hit rates for free. GPT-5.6 dropped partial-prefix matching:
I think for a company in AI specifically it's worse.
It makes me feel like either
1) you don't use the models enough to know how they write
2) you're not self aware enough to know it matters
3) you're oblivious to the situation overall
4) you don't respect your readers
There's no good scenario.
Can we get over the detective work about if the text was written by LLM or not in 2026 already ? This is a lost cause, and we could instead focus on substance over syntax.
Not OP but my frustrations come from it being impossible to ignore and outright distracting.
I've found the same thing showing with Claude-coded/designed front ends that overuse the same semi-monospaced fonts, Blue/Yellow/Red palette and rounded corner borders. It isn't that it is bad, but it often isn't fit for purpose.
You're right it wont change anything, but authors shouldn't be surprised when people who care about their time/attention comment on low/no effort pieces.
critique of writing style isn't made better by claiming it was authored by an LLM
No but it's a useful shorthand to describe a type of bad writing.
I also think that people should focus on substance and not if AI was used, but AI writes like shit and I find myself retching a bit when I have to read long AI-written documents. Do they say something useful? Maybe, but when my eyes are glazing over because it's just so exhausting trying to parse what's written, I can't tell.
I certainly think less of people when they have such poor taste that they think writing like that is acceptable.
This is just an AI comment masquerading as not trying to prove a point.
Since you didn't address the substance, I guess we'll never know :)
I’m as pro AI as any.
Slop is slop. When the “it works but isn’t great” phrases end up slipping into a strong conceptual core, it compromises the perception of the ideas.
Perhaps our AI will cater to us by rewriting the content we read, and each of us intermediate all communication with systems that make that slop bearable.
Or perhaps, we learn that we kinda still need to give a shit when writing to land on the perception we’re trying to create within our readers
To me it's a useful signal not to read an article that someone didn't bother to write.
Which is a shame as real insights are buried inside some of these articles, which if the author bothered to write in his own words could have reached an audience that would have appreciated them.
Writing is one of the areas where I want no LLM involvement.
More often it's the difference between finishing something and not finishing it - so often LLM's are helping them reach an audience that would appreciate them, even if that audience doesn't include you
The problem is that the second you suspect something is written by AI, its a pretty good signal that 50-80% of the text is empty of meaning. Maybe that will change, but LLMs are terrible and inefficient writers.
Only so much time in the day, its a quick signal to not waste anymore of it.
Correct. AI == Credibility hit and it's increasing as more humans get used to feeling they are AI slop consumers, not worth the time for genuine human engagement. Human engagement costs are increasing. Amazing to read/watch.
It's not about figuring out if it's LLM written though. The style is hard to read and annoying. With the kind of sentences GP was talking about it's actually harder to get the substance.
It means I don’t trust the substance. Whenever I try to use for technical writing like this, I catch it getting things wrong constantly.
Yes, as soon as models come out that can write properly, we'll all instantly get over it. Until then we'll be having this discussion over and over, as many times as it is necessary.
It's both poor substance and style, in most cases (and this case.)
Pointing out they generated it at least encourages them to write a shorter article that says what they meant.
No. As a human I like reading human written text over computer written text. I want something a human composed with thought put into it. Not something a human tried to save time with by having the machine write it.
the frustration is largely because the overall substance is quite poor since it is typically imprecise by nature.
Evaluating substance takes time - perhaps more than was invested in the article to begin with. So these tells are very distracting because as soon as I see them I wonder if the person who prompted the LLM even bothered to read the output. If they haven't, then I certainly shouldn't invest the time to determine if there is any substance.
What substance? That they consume a newer model from the same vendor?
Because quality of writing matters.
Good communicators learn to use the written word. Bad ones rely on mental crutches.
Good communicators get an audience, and bad ones won't.
You think it's a lost cause, but it's not, because people don't like this junk, because it is low quality and, on average, lacks substance.
The best minds in AI that I've seen all write their own words. They use AI to help them research or ideate, but what they write is their own.
Before assuming this is a "lost cause," consider why the smartest people in the room don't do it.
The syntax tells you there isn't any substance.
The substance is shit too with these LLM articles. Stuck in the box of the training set. Nothing new. Just regurgitation.
I have a counter proposition: don't fall for this constant suggestion that LLMs are an unavoidable future would you leave the techbros alone now pretty please, relentlessly keep reminding that we still don't think it's acceptable so people don't start to think this is okay since nobody complains anymore.
I appreciate these comments, they save me time for procrastinating elsewhere.
I agree with this sentiment: it’s not inevitable if we relentlessly ostracize obviously LLM posts
And let’s be real: I had a post this year that was #1 on HN for a while, and an LLM “wrote” the whole thing, but it was very much my writing style and NO ONE called out the post as LLM slop. If you use an LLM correctly for writing, it’s not detectable. It seems that most folks don’t go through that effort.
Makes you wonder if any of stats these articles push are even real.
You should make sure to not read Stratechery then. It's writing is even worse.
Gets a 100% on Pangram. Stuff is so distracting. Write your own posts, FFS. Or at least pass it through "humanizer" type plugins.
> The way the LLMs write (Claude perhaps?) With short phrases separated by colons, commas or full stops, is so poor and frustrating.
Yup llmish (from now on it's called "llmish") sucks.
But I'd say: at this point it's probably trivial to write a browser extension that detects llmish and that rewrites the worst sentences: from llmish to something less irritating to read. Heck, I could spent tokens on that: an extension that changes on the fly llmish found on webpages.
Also I'd say there's typically no swearing at all in llmish: llmish is too politically correct for swearing. So the rewrite could maybe also use a few "offending" words.
Offending words that, btw, are not going to go well with Gen Zers. Poor Gen Z... They've been raised with the state and its institutions (like school and then universities) hammering them with the notion that they were precious little unique snowflakes and now they arrive on the job market only to be told they've been pre-emptively replaced by AIs. And because they cannot stand a single curse word (because it's "offensive to minorities" or something), they'll be driven off by text rewritten to contain curse words. So they're condemned to read the bland, dumb, AI-generated llmish for the rest of their lives.
Honestly sucks for them. Fuck that.
Have you ever met an actual Gen Z? They have no problem with swear words. Many of them love Key and Peele, whose humor is like 90% racist jokes.
If wokeness actually did capture a whole generation then why even bother complaining?
For me, it's Bottish
We run a lot of varied, tiny, simple workflows that were previously running on 5.4-nano and mini. We transitioned them to 5.6 and noticed exactly this range of improvement across the board. In a few cases, we had improvements in classification.
I think a lot of people miss that for many companies, a model upgrade like this is basically a one liner.
Even if you have an amazing model router architecture (which we do for our golden flows), it’s just not worth it. Not to mention reliability and so on
> Ploy’s agent builds and edits real marketing websites. It plans a page, reads the codebase, writes components, generates imagery, screenshots its own work, and decides when it’s done. That job description sets a very high bar for a model, and we test every frontier release against it. For the four months Opus held the default slot (first Opus 4.7, then 4.8), nothing we tested beat it.
Well, unlike OP I haven't run a rigorous test, but I still would expect Fable to be significantly better at building marketing websites than Opus. It sure is way better at building decks.
4.7 is very autistic in terms of following directions so I find OPs claims plausible
Very descriptive there heh
Game recognises game
Migrating my workflow to Reasonix with cache hits on Deepseek make requests practically free, and that's on unsubsidized American providers.
Sorry, what did that have to do with the article?
They also migrated and that also made the workflow cheaper.
It has everything to do with the article.
The cost reduction is impressive, but I think consistency matters even more for production agents. I'd be interested to know whether prompt engineering or tool-calling workflows had to change significantly.
> we’ve made GPT 5.6 Sol the default model powering every Ploy workspace
I would consider Luna for parts of the workload that touch actual tools. It is surprisingly capable and it runs fast.
Sol is great at talking to the human and orchestration of agent calls, but it's just too expensive to use everywhere.
You can get 5 Luna runs for the cost of 1 Sol run. Statistically speaking, going from one to five samples is a pretty big deal.
The problem I always run into with subagents is that they are isolated. This is a double-edged sword, as it keeps context down and lets them "focus", but it often means they must do their own research to continue to do work given to them, which eats uncached tokens.
So depending on how heavily agents are used on what tasks, it's entirely possible that you get worse work for more cost.
> they are isolated
This is a feature if your goal is to obtain many samples. Independence is critical. This makes it easier to accurately model the uncertainty of a decision.
I feel Claude Code has added (and removed?) a feature that forks a subagent from the parent context, so it’s still isolated but it’s more of a continuation of what you were doing in one narrow direction and then it dies. Rather than a blank slate with a prompt of what to do.
But what users prefer? Given this is for marketing, which results produce more conversions? From the examples shown, personally I strongly preferred Claude Opus in all cases.
Thank you for a dense informative article with practical takeaways. This was an easy read and it reinforced the importance of some concepts in LLM based pipeline design.