Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper(ploy.ai)
ploy.ai
Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper
https://ploy.ai/blog/migrating-a-production-ai-agent-to-gpt-5-6
98 comments
You are reading it wrong. Ask your LLM to read and summarize based on your style preferences. Better yet don’t read anything at all, just tell your agent to convert it to a skill file for it’s future reference..
Unironically, I think in the future we will have the option to run filters in our browsers that can reword articles to your preferred writing style. Like user stylesheets for text. I also assume it already exists out there, I've just been to lazy so far to look for it. It's a relatively obvious application of LLMs.
I've used that for about 15 years because I live abroad, it's called Google translate.
this is probably sarcasm, but might actually try this. the visceral negative reaction i have to llm writing makes me instantly want to close the tab
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.
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.
Whenever I suspect an article is LLM authored I stop reading it immediately and instead give it to my LLM tool of choice to summarize/ paraphrase.
This way I can at least somewhat control the style of the output.
This way I can at least somewhat control the style of the output.
Makes you wonder if any of stats these articles push are even real.
I would rather read an article with actual production experience migrating an agent, even if it is written in this style, than a perfectly crafted long read from another evangelist that has nothing but high level fluff and general phrases about the bright future of ai
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.
My solution to this is to dump it into an LLM and a prompt that roughly does something like...
Hand wavy list just to get a general idea...
1) Give me a condensed summary 2) Is this adding anything to what we already have? (I save good articles along with annotations and whatever notes I may write to go along with it.) 3) Locate any upstream ideas on this (often AI articles are rehashing much better written ideas.) ...
Something like that. Not that I have some great system for it. I find these articles are so full of fluff that I have lost patience to attempt to get through them. So, I pull out the AI to parse the AI. I know that the AI may miss some hidden gems, but I'm okay with that.
Hand wavy list just to get a general idea...
1) Give me a condensed summary 2) Is this adding anything to what we already have? (I save good articles along with annotations and whatever notes I may write to go along with it.) 3) Locate any upstream ideas on this (often AI articles are rehashing much better written ideas.) ...
Something like that. Not that I have some great system for it. I find these articles are so full of fluff that I have lost patience to attempt to get through them. So, I pull out the AI to parse the AI. I know that the AI may miss some hidden gems, but I'm okay with that.
> 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.
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.
As a certified Gen Z member: respectfully, what the fuck are you on about?
The mythical Gen Z you are describing is the hyperbolic exaggeration that is equivalent to me describing all boomers as racist, all Gen X as entitled, all millennials as lazy, etc.
The mythical Gen Z you are describing is the hyperbolic exaggeration that is equivalent to me describing all boomers as racist, all Gen X as entitled, all millennials as lazy, etc.
For me, it's Bottish
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?
If wokeness actually did capture a whole generation then why even bother complaining?
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.
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.
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.
To be fair internet (or rather shovelware websites like Medium) were already flooded by crap articles based on a set of templates. Of course the issue is that LLMs are actually better than the robot-humans who used to write them so now it takes more time until you figure whether it’s worth reading or not..
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
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
Maybe if instead we stopped engaging they would wonder why
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.
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.
I keep coming back to this https://tombedor.dev/human-attention-and-human-effort/
The number of things that make it to the top of HN/Reddit/wherever now that are devoid a human's touch is exhausting. Whether it's a site that's got that Claude frontend smell, or a repo that's got a burst of 10 claude commits before getting shared and abandoned, or a series of blog posts that were written by LLMs... it's all, at this point, a flag for me that the human behind the LLM doesn't really want to engage with others or share; in some ways it dehumanises their entire (supposed) audience.
IDK. Maybe having Claude contribute writing about something novel to the general blogosphere is useful in some dimension, but it usually gives me no confidence in the truth of the post.
The number of things that make it to the top of HN/Reddit/wherever now that are devoid a human's touch is exhausting. Whether it's a site that's got that Claude frontend smell, or a repo that's got a burst of 10 claude commits before getting shared and abandoned, or a series of blog posts that were written by LLMs... it's all, at this point, a flag for me that the human behind the LLM doesn't really want to engage with others or share; in some ways it dehumanises their entire (supposed) audience.
IDK. Maybe having Claude contribute writing about something novel to the general blogosphere is useful in some dimension, but it usually gives me no confidence in the truth of the post.
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
I agree with the copy-pasted slop that you see sometimes, that is probably generated from a short prompt and therefore has no real substance to it.
If an article has interesting content (which comes from an human) and the LLM is just used to help the author finish off the article, I don't have any problem with that.
Labeling both scenarios in the same category feels completely wrong to it, as equating vibe coded stuff (as in no human ever read the produced code) and agent-assisted good old software engineering
If an article has interesting content (which comes from an human) and the LLM is just used to help the author finish off the article, I don't have any problem with that.
Labeling both scenarios in the same category feels completely wrong to it, as equating vibe coded stuff (as in no human ever read the produced code) and agent-assisted good old software engineering
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.
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.
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.
Pointing out they generated it at least encourages them to write a shorter article that says what they meant.
Its never lupus and its always an LLM. No need for detective work.
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.
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.
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.
What substance? That they consume a newer model from the same vendor?
the frustration is largely because the overall substance is quite poor since it is typically imprecise by nature.
100% this.
The AI police is there to say what is worth reading and what is not, because THEY know what people like.
Or not.
The AI police is there to say what is worth reading and what is not, because THEY know what people like.
Or not.
It's not detective work, it is literally blatantly obvious and impossible to ignore.
And no we can't get over it either. But I already have talked about that before and said roughly all I have to say on that front, so I'm just going to link back to my last comment regarding this.
https://news.ycombinator.com/item?id=48861849
And no we can't get over it either. But I already have talked about that before and said roughly all I have to say on that front, so I'm just going to link back to my last comment regarding this.
https://news.ycombinator.com/item?id=48861849
But it's not about whether it was AI or human authored. That misses the point. We're all fatigued on the writing style. The same cadence and patterns; the same phrases and terms like "load-bearing". Used everywhere, they create a super fatiguing monoculture in all the writing. It's like if every illustration on the internet suddenly contained Garfield the cat.
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.
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.
No.
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 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.
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.
As of today, Ploy’s agent runs on GPT-5.6 Sol, the flagship tier of the model family OpenAI released this morning.
Its ironic that under an article with a ton of deep infrastructure insights half the comments are crying about the "forced writing style". What does it matter if claude helped the author clean up the text when inside is a ready-to-use blueprint on how to save 30% of the api budget and fix empty file reads?
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
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
the article is literally about the model upgrade not being a one liner
What SDK are you using? Or is it custom?
The first thing I used Sol for was to assess 5.6 on our workflows - previously, it was 5.5 for everything, as the quality on simpler models was just not good enough. We’re doing a mix of text and image analysis to extract explicit and implicit structured data from a steaming pile.
They do work pretty much as advertised. The bulk of our workload is now going via terra, which has cut our cost in half by itself, as well as improved response times by 50% - luna I am using as a backstop for opencv hits, and it is good enough, and so cheap as to almost be free - but very limited - and very fast. Sol only gave marginal improvements over terra for our workload.
I’ve also gotta say I’m impressed as to how well Sol ultra carried out the assessment itself - it made sound recommendations, and gave me a nice big dossier of “you should look at these outputs yourself and compare and consider” along with raw and digested data, and cpm for queries.
Anyway. Spent nothing beyond my pro sub, let Sol gnaw on it for a few hours, and my cost basis just dropped 50% and throughput improved by 100%. Win.
They do work pretty much as advertised. The bulk of our workload is now going via terra, which has cut our cost in half by itself, as well as improved response times by 50% - luna I am using as a backstop for opencv hits, and it is good enough, and so cheap as to almost be free - but very limited - and very fast. Sol only gave marginal improvements over terra for our workload.
I’ve also gotta say I’m impressed as to how well Sol ultra carried out the assessment itself - it made sound recommendations, and gave me a nice big dossier of “you should look at these outputs yourself and compare and consider” along with raw and digested data, and cpm for queries.
Anyway. Spent nothing beyond my pro sub, let Sol gnaw on it for a few hours, and my cost basis just dropped 50% and throughput improved by 100%. Win.
My experience mirrors this: services like OpenRouter that promise “failover” are pretty much useless except for sandbox testing because models in production are not really interchangeable. Any production harness doing serious agentic work in production is dependent on more model-specific quirks than you would expect. And even if another model works without errors, performance and efficiency is a whole different story. Even the system prompt can and should be tuned to a model’s preferred speaking style, for example <xml tags> for Claude-like models because they were trained on it, while other models do better with other delimiters. Think of the whole harness, prompt, and model as one system, not really with modular parts that can be swapped out if you care about optimal performance.
I believe part of the LLMOps (I don't like the term, but it is what it is) should be building a failover plan with proper testing that check tools trajectories and such. If you have these then you can sort the good enough models from cheaper to more expensive and have the failover you mentioned.
I saw people bulding a mapping of model->{{prompts}, {tools descriptions}, ...}, but that, to me, it feels extreme. I believe it is the model that needs to adapt to your prompts after a certain point. Models that fail to do so won't get our api requests as they will be out of the failoever roster.
I saw people bulding a mapping of model->{{prompts}, {tools descriptions}, ...}, but that, to me, it feels extreme. I believe it is the model that needs to adapt to your prompts after a certain point. Models that fail to do so won't get our api requests as they will be out of the failoever roster.
OpenRouter doesn't fail over to a different model, it fails over a different provider of the same model.
> 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.
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.
gpt 5.6 is so much better ar design than fable
4.7 is very autistic in terms of following directions so I find OPs claims plausible
> The fix that worked is a schema transform at the provider boundary. For OpenAI-family models only, we rewrite every optional property to be required but nullable, using anyOf: [T, null], which gives the model an explicit way to say “not using this.”
I admit, I've only used a bastardized form of MCP, but this smells... wrong? It's not clear to me why the Typescript type definitions would have any influence on (what I presume is) JSONSchema being sent from the agent to the inference backend as part of the completion request. The MCP specification (which the OpenAI backend might not use, I don't know) has an explicit field to signify "optional" parameters in the JSONSchema; my read on this is there's a bug somewhere between the Typescript layer(??) and the generated tool description which is actually sent to the inference backend.
It's possible the inference backend has changed from "generate valid tool responses" to "generate valid tool responses according to the JSON schema [where no parameters are optional]" but it's impossible to tell without seeing the actual requests sent to the inference backend (which I didn't see in TFA).
I admit, I've only used a bastardized form of MCP, but this smells... wrong? It's not clear to me why the Typescript type definitions would have any influence on (what I presume is) JSONSchema being sent from the agent to the inference backend as part of the completion request. The MCP specification (which the OpenAI backend might not use, I don't know) has an explicit field to signify "optional" parameters in the JSONSchema; my read on this is there's a bug somewhere between the Typescript layer(??) and the generated tool description which is actually sent to the inference backend.
It's possible the inference backend has changed from "generate valid tool responses" to "generate valid tool responses according to the JSON schema [where no parameters are optional]" but it's impossible to tell without seeing the actual requests sent to the inference backend (which I didn't see in TFA).
The thing is this isn't a schema generation or Typescript bug at all. This is just how openai's function calling works under the hood. Their weights were fine-tuned for tool use to output the most complete data structures possible. If the model sees a parameter name in the system prompt context it will try to fill it with a value, even if it is not in the required array
Modern frontier models, including Fable/Opus and 5.6, are often very loose with tool calls, and often don’t follow your schema precisely.
For example, see this post for Claude models hallucinating properties for an edit/replace tool call in Pi: https://lucumr.pocoo.org/about/
I suspect some part of this comes from the noticed intelligence degradation when you do constrained decoding. Yes, you’re guaranteed schema validation, but you lose a lot of intelligence. It’s fine if you just want a classifier, a summary, a prompt enhancement, etc; but I’d be careful in agentic loops.
Harnesses like Claude Code do a lot of preprocessing, repairing, cleaning, etc; as the blog post shows. You usually don’t see it.
In practice it’s easier and better to just make your harness “looser” and work better with the model (they’re coming out every month or two anyway, each with their own idiosyncrasies) than to assume and force perfect correctness.
Welcome to vibe applied AI ;)
For example, see this post for Claude models hallucinating properties for an edit/replace tool call in Pi: https://lucumr.pocoo.org/about/
I suspect some part of this comes from the noticed intelligence degradation when you do constrained decoding. Yes, you’re guaranteed schema validation, but you lose a lot of intelligence. It’s fine if you just want a classifier, a summary, a prompt enhancement, etc; but I’d be careful in agentic loops.
Harnesses like Claude Code do a lot of preprocessing, repairing, cleaning, etc; as the blog post shows. You usually don’t see it.
In practice it’s easier and better to just make your harness “looser” and work better with the model (they’re coming out every month or two anyway, each with their own idiosyncrasies) than to assume and force perfect correctness.
Welcome to vibe applied AI ;)
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.
It has everything to do with the article.
What's your config, how does it compare to pi
> 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.
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.
Statistically speaking if each part of the Luna run has a 90% chance of being correct, 5 of those is 0.9^5 = 0.59 = 59%. Or one Sol run being maybe 95% correct? Exact numbers vary of course. But then again having sol verify at end may be cheaper.
The goal is not 100% correctness. The goal is to demonstrate the current amount of variance / uncertainty to the planning agent.
If the planner sees that 4/5 Luna runs resulted in approximately the same summary, it may conclude that variance is low and that it is over the target. If all Luna runs are different, the planner can conclude that additional research rounds are required.
If the planner sees that 4/5 Luna runs resulted in approximately the same summary, it may conclude that variance is low and that it is over the target. If all Luna runs are different, the planner can conclude that additional research rounds are required.
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.
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.
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.
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.
He actually talks about that in the article. Including how they had to rearchitect tool calling using nulls as well as limitations of prompt caching
Personally, could you share the code sometime later? The GPT code looks decent to me. Or even just the prompt would be fine.
I catch a lot of issues on the technical writing side.
I found Claude to be better for the first prototype. It was more likely to come up with something fast. But it kept lying and claiming it did world class work and it was just hardcoding response by the end. I found GPT never lied to me.
We at Playcode.io - a company similar to Ploy are still using Opus 4.6. "Why?" you might ask.
Because GPT 5.6 Sol, while fast and pleasant to use, is essentially the same model as 5.5 wrapped in new marketing packaging, just to avoid losing ground to Anthropic. In practice, it's the same quality: it generates the same garbage, tons of code, and can never solve even a single complex task. We simply don't trust it to write code for clients that they'll end up throwing away anyway.
"Then why not Opus 4.8?" you might ask.
Well, because Opus 4.8 and 4.7 are just another lie, a price hike with no actual quality improvement.
That's why at Playcode, we give our clients the best possible quality/price - which is Opus 4.6. Regardless of what people write in articles like this.
Because GPT 5.6 Sol, while fast and pleasant to use, is essentially the same model as 5.5 wrapped in new marketing packaging, just to avoid losing ground to Anthropic. In practice, it's the same quality: it generates the same garbage, tons of code, and can never solve even a single complex task. We simply don't trust it to write code for clients that they'll end up throwing away anyway.
"Then why not Opus 4.8?" you might ask.
Well, because Opus 4.8 and 4.7 are just another lie, a price hike with no actual quality improvement.
That's why at Playcode, we give our clients the best possible quality/price - which is Opus 4.6. Regardless of what people write in articles like this.
Everyone probably has the same question: what about Fable? Fable 5 is sick.
It is simply the best model in the world out of everything we have ever tried. It's absolutely fantastic. It solves almost any task from start to finish, the way it should be done — no errors, perfect code. It's a miracle.
If there's any way to make it a little more affordable, that would be incredible.
As for GPT-5.6 Sol — it doesn't even come close. I honestly don't understand why people even try to compare them. It feels like Sam's attempt to hold onto his audience with those endless daily limit resets. A clever trick, nothing more.
It is simply the best model in the world out of everything we have ever tried. It's absolutely fantastic. It solves almost any task from start to finish, the way it should be done — no errors, perfect code. It's a miracle.
If there's any way to make it a little more affordable, that would be incredible.
As for GPT-5.6 Sol — it doesn't even come close. I honestly don't understand why people even try to compare them. It feels like Sam's attempt to hold onto his audience with those endless daily limit resets. A clever trick, nothing more.
> Fable 5 is sick. [It] solves almost any task from start to finish, the way it should be done — no errors, perfect code. It's a miracle.
> As for GPT-5.6 Sol — it doesn't even come close. I honestly don't understand why people even try to compare them.
What kind of problems are you working on? I like Fable but when planning work on a complex C codebase it's making more mistakes than 5.6 Sol xhigh for me.
In what scenarios is Fable giving you "no errors, perfect code"?
> As for GPT-5.6 Sol — it doesn't even come close. I honestly don't understand why people even try to compare them.
What kind of problems are you working on? I like Fable but when planning work on a complex C codebase it's making more mistakes than 5.6 Sol xhigh for me.
In what scenarios is Fable giving you "no errors, perfect code"?
For PCB design 5.6 Sol/Terra is streets ahead of 5.5, and uses fewer tokens, so I'm not sure it can really be the same model.
Did you just paste your marketing copy into a hn comment?
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.
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: