I think Warhol’s quote is nostalgic but incomplete.
I’m priced out of the best cars, best houses, best home theater systems, best schools. Even someone making $300k/year can’t afford all of the best of everything.
Sure, the iPhone has been “the best” possible phone which was also used by nearly everyone, but I think that’s an anomaly even in the short run.
Right now I’m paying $200/mo for Claude code to do an amount of work I would’ve had to pay $10,000/mo for. Of course I’m expecting those numbers to get closer to each other.
If you know anything about tech, you will know that tech as an industry is highly deflationary--billionares use the same iPhones as you do! (in contrast, they don't drive the same cars you do)
This boils down to the fact that chip fabs have massive fixed costs and near-zero marginal costs, and these chips power all of tech. So the more chips they can produce for a given fab, the more profit they can make, meaning that companies are incentivized to sell as many products as possible for as low a price as possible.
We're supply constrained in the short-term because demand for these AI tools is so high that TSMC and other chip manufacturers can't keep up. But long term, supply/demand will equalize and tech will continue its deflationary trend. Sure, the frontier will always require the best possible chips, but AI coding is highly competitive, and competition drives price decreases. So prices may stay high right now, but it seems unlikely to me that this will stay true long-term.
All four of the author's steelmanned arguments at the end for a price decrease seem likely to come true already: competition is intense (OAI brags about how much cheaper they are compared to Claude), OAI subsidizes open-source influencers already, companies' earnings calls all call for more investment in fabs, and we're already close to saturating all of the benchmarks used for RL!
"OpenAI reportedly discussed charging $20k/month on PhD-level research agents with investors."
I've been wondering about this, that there might be a day when certain models are sold at a much higher price, like luxury cars, and only people who are willing to pay a lot of money get them. Everyone else has to settle for a cheaper LLM.
> The top tier subscription prices are increasing exponentially
WILD graph that misrepresents what is happening.
There's a bunch of $20 subscriptions, and a bunch of $200 subscriptions. Devin has a $500 subscription. That's it.
The cost per unit of intelligence has been dropping every month. The cost per "completed task" has also been dropping. There is no sign of this reversing course. Graphing the price of a subscription, without taking into account what that subscription is getting you, is poor authorship.
Imagine if a model ever does get scary good, would the big labs even release it for general use? You couldn't even buy it if you wanted to. Exceptions would be enterprise deals / e.g.: $AMZN niche super contracts.
Companies have built entire systems of such complexity and slop that they require AI just to do the maintenance. They have fired engineers thinking they can just replace them with AI.
Well when the prices rise, they have no choice but to stay locked in, paying whatever it takes just to keep their companies running. If they stop using AI, their workforce suddenly does not have capacity to do the work required because of the layoffs. And there are not enough people to hire because people are quickly turning away from software engineering as a career. What a disaster it will be.
> the cheapest usable tier of Claude Code is $100/mo
is, imo, false. cc pro, $20 per month, gets you a lot of sonnet usage, and code review with opus (which i find very valuable, even as someone who tries to use ai little). i guess it depends how you use ai, but if you use it to plan, debug, and review, rather than having it write code, i think pro is pretty comfortable.
to add, i've seen people say these subscriptions will get far more expensive, as they're offered at a loss. but, it seems far more likely that free tiers will be degraded or disappear, as (especially for openai?) the relative number of subscribers to free users is very small, so the latter probably dominates compute time greatly. anthropic probably has a higher relative number of people who pay for claude code (and use it to its fullest), so this is probably less true. i can see pro getting less usage, and max increasing in cost.
AI providers can only charge what the market can bear. AI isn't worth 20k/month for 'PHD' level work. But people are willing to pay for several $200/month subscriptions.
But fundamentally AI compute is a commodity. GPUs are made in factories at scale. Assuming AI quality tapers off eventually supply will catch up to demand.
Finally open weights models are good enough that the leading labs cannot charge high margins.
> OpenAI reportedly discussed charging $20k/month on PhD-level research agents with investors.
At this price point, it will be cheaper to hire a bunch of actual PhDs. The vast majority who will not earn anything close to 250k per year in most of the world.
Yes, but I suspect part of the pitch is that the “PhD level models” will accomplish 50x as much in a year as the human PhD, because they’re faster AND they work 24/7.
Whether they can deliver is another question, but I wouldn’t bet my career that they can’t.
From my recent experience with Qwen 3.5 I am less concerned about this. It certainly will never be “the best” but I did some TS refactoring with Qwen + Opencode over the weekend and it was surprisingly good. I even asked Opus 4.6 to grade the commits and it usually gave it a B- haha..
Anyway, it might be worth it to invest in an LLM rig today if you’re paranoid.
I've been thinking about this as well, and I'm glad the author is talking about it. However, I don't think he took it far enough.
It is correct to say there's near-infinite demand for AI, and supply is limited. It stands to reason that wealthier people will pay more, and therefore get more, out of AI.
However, this has always been true, but historically instead of AI it's been workers. The economics of labor haven't changed. So it will, as always, be a game of how you deploy the workers you hire. Are you generating useless morning briefs or are you actually generating value for yourself and others with the AI you buy? If you generate more value that the tokens you burn, you'll get ahead.
This will be true in academia as well, the area of interest to the author. He writes like, before AI, grad student level intelligence came for free.
The only way this happens is if models that are specifically made to do certain kinds of coding start to exist. Then this would start to become an issue, yes, until those models are distilled into smaller models.
I don't agree. There are a lot of inference performance improvements to make. I think the cost of inference continues to fall, and pretty much every application of AI becomes a commodity with brutal competition.
they don't theoretically have to aside from that industry going that direction and the stickiness of communication through it, but it simplifies some of their job
it didn't revolutionize trading or make it more democratized, despite simplifying some aspects of the industry
the technology could have but it remains a specialized tool
thats the way I see agentic coding tools and the trend is following it
once the UX designers, PMs and ideas guys get bored of their newfound SaaS slop capabilities, it will be back to specialists doing this and nobody else
Regardless if the exponential trend the author writes about is correct or not, I do think the cost of AI is reversing the trend for coding. For quite some time the tools we used have become cheaper and cheaper and more avaliable than before. Nowadays compilers, IDEs and other tools are increasingly open source or at least free or very cheap. But with AI its no longer the case.
"The underlying purpose of AI is to allow wealth to access skill while removing from the skilled the ability to access wealth. --@jeffowski"
While I don't think that's the only purpose, I can't help but think that people that become dependent on these tools will have neither wealth nor skill. Keep your skills sharp!
I predicted months ago that $20/month isn't going to fly anymore. I think if it produces code, it will jump to $1000/month at least. The value of the LLM nowadays is much higher than $1000/month and I think we will see that happen in 2026, because these companies need to make money ASAP in order to get more funding for the next round of training.
I agree with the core assumption (to a point -- there is a point of diminishing return where pretty excellent tools are cheap), but this line is ridiculous-
"the cheapest usable tier of Claude Code is $100/mo"
Bullshit. I have the $20 plan and seldom hit the quota. I used to hit the distinct Opus quota, but now that isn't separate I just don't anymore. Even enabled the extra quota charging and have never paid a penny more.
And to be clear, to most people I'm a pretty heavy user. Like, practically it has a heavy influence on my day to day work, and is an amazing contributor to my functions.
The people who think only the $100+ tier is "usable" are often (albeit not always) usually the people doing the worthless, but "forward-thinking" nonsense, throwing millions of tokens aspirational. Like the OpenClaw nonsense is 99.99% worthless filler where people chased a productivity hack that in reality is just hobbyist silliness.
It's token shredding for almost no value as people show that they're with it. People gloating about their swarms of agents doing effectively nothing are another "I need to max out everything" people. These are the ones who yield the result that AI has no benefit to productivity, as they overdo something to such ridiculous extremes.
The same for the laughably poorly thought out MCP servers that flood a service with a quarter million tokens for negligible value. So much insanely poorly considered nonsense is in use, to the great glee of the AI companies. And, I mean, I guess I should thank these people for basically subsidizing it for the rest of us.
The rest of us are surgically applying AI precisely, to incredible effect. The cheap plans are ridiculously valuable.
37 comments
[ 3.3 ms ] story [ 40.3 ms ] threadI’m priced out of the best cars, best houses, best home theater systems, best schools. Even someone making $300k/year can’t afford all of the best of everything.
Sure, the iPhone has been “the best” possible phone which was also used by nearly everyone, but I think that’s an anomaly even in the short run.
Right now I’m paying $200/mo for Claude code to do an amount of work I would’ve had to pay $10,000/mo for. Of course I’m expecting those numbers to get closer to each other.
No VC-funded gravy train lasts forever.
So maybe it's true you won't get "the best", but I don't think you'll be that far off.
This boils down to the fact that chip fabs have massive fixed costs and near-zero marginal costs, and these chips power all of tech. So the more chips they can produce for a given fab, the more profit they can make, meaning that companies are incentivized to sell as many products as possible for as low a price as possible.
We're supply constrained in the short-term because demand for these AI tools is so high that TSMC and other chip manufacturers can't keep up. But long term, supply/demand will equalize and tech will continue its deflationary trend. Sure, the frontier will always require the best possible chips, but AI coding is highly competitive, and competition drives price decreases. So prices may stay high right now, but it seems unlikely to me that this will stay true long-term.
All four of the author's steelmanned arguments at the end for a price decrease seem likely to come true already: competition is intense (OAI brags about how much cheaper they are compared to Claude), OAI subsidizes open-source influencers already, companies' earnings calls all call for more investment in fabs, and we're already close to saturating all of the benchmarks used for RL!
I've been wondering about this, that there might be a day when certain models are sold at a much higher price, like luxury cars, and only people who are willing to pay a lot of money get them. Everyone else has to settle for a cheaper LLM.
WILD graph that misrepresents what is happening.
There's a bunch of $20 subscriptions, and a bunch of $200 subscriptions. Devin has a $500 subscription. That's it.
The cost per unit of intelligence has been dropping every month. The cost per "completed task" has also been dropping. There is no sign of this reversing course. Graphing the price of a subscription, without taking into account what that subscription is getting you, is poor authorship.
Imagine if a model ever does get scary good, would the big labs even release it for general use? You couldn't even buy it if you wanted to. Exceptions would be enterprise deals / e.g.: $AMZN niche super contracts.
Companies have built entire systems of such complexity and slop that they require AI just to do the maintenance. They have fired engineers thinking they can just replace them with AI.
Well when the prices rise, they have no choice but to stay locked in, paying whatever it takes just to keep their companies running. If they stop using AI, their workforce suddenly does not have capacity to do the work required because of the layoffs. And there are not enough people to hire because people are quickly turning away from software engineering as a career. What a disaster it will be.
> the cheapest usable tier of Claude Code is $100/mo
is, imo, false. cc pro, $20 per month, gets you a lot of sonnet usage, and code review with opus (which i find very valuable, even as someone who tries to use ai little). i guess it depends how you use ai, but if you use it to plan, debug, and review, rather than having it write code, i think pro is pretty comfortable.
to add, i've seen people say these subscriptions will get far more expensive, as they're offered at a loss. but, it seems far more likely that free tiers will be degraded or disappear, as (especially for openai?) the relative number of subscribers to free users is very small, so the latter probably dominates compute time greatly. anthropic probably has a higher relative number of people who pay for claude code (and use it to its fullest), so this is probably less true. i can see pro getting less usage, and max increasing in cost.
But fundamentally AI compute is a commodity. GPUs are made in factories at scale. Assuming AI quality tapers off eventually supply will catch up to demand.
Finally open weights models are good enough that the leading labs cannot charge high margins.
At this price point, it will be cheaper to hire a bunch of actual PhDs. The vast majority who will not earn anything close to 250k per year in most of the world.
Whether they can deliver is another question, but I wouldn’t bet my career that they can’t.
Businesses will pay more since they can justify the cost. That seems fine?
Unless you’re a top-tier domain expert. Then you’re safe until (if…) ASI.
Anyway, it might be worth it to invest in an LLM rig today if you’re paranoid.
Everything points to commoditization of models. Open/distilled models lag behind frontier only by 6-12 months.
Regulatory capture is the only thing I’m scared of with regards to tooling options and cost.
It is correct to say there's near-infinite demand for AI, and supply is limited. It stands to reason that wealthier people will pay more, and therefore get more, out of AI.
However, this has always been true, but historically instead of AI it's been workers. The economics of labor haven't changed. So it will, as always, be a game of how you deploy the workers you hire. Are you generating useless morning briefs or are you actually generating value for yourself and others with the AI you buy? If you generate more value that the tokens you burn, you'll get ahead.
This will be true in academia as well, the area of interest to the author. He writes like, before AI, grad student level intelligence came for free.
Ok, wait, sorry, bad example...
"Let's just make random shit up and expand it into a whole blog post."
Seriously does anyone believe this premise? The Claude Max ($200/mo) is the same kind of product as Github Copilot ($10/mo) so the price 20x-ed?
they don't theoretically have to aside from that industry going that direction and the stickiness of communication through it, but it simplifies some of their job
it didn't revolutionize trading or make it more democratized, despite simplifying some aspects of the industry
the technology could have but it remains a specialized tool
thats the way I see agentic coding tools and the trend is following it
once the UX designers, PMs and ideas guys get bored of their newfound SaaS slop capabilities, it will be back to specialists doing this and nobody else
I wrote more about this in a blog, at https://www.viblo.se/posts/ai-hobbycoding/
"The underlying purpose of AI is to allow wealth to access skill while removing from the skilled the ability to access wealth. --@jeffowski"
While I don't think that's the only purpose, I can't help but think that people that become dependent on these tools will have neither wealth nor skill. Keep your skills sharp!
"the cheapest usable tier of Claude Code is $100/mo"
Bullshit. I have the $20 plan and seldom hit the quota. I used to hit the distinct Opus quota, but now that isn't separate I just don't anymore. Even enabled the extra quota charging and have never paid a penny more.
And to be clear, to most people I'm a pretty heavy user. Like, practically it has a heavy influence on my day to day work, and is an amazing contributor to my functions.
The people who think only the $100+ tier is "usable" are often (albeit not always) usually the people doing the worthless, but "forward-thinking" nonsense, throwing millions of tokens aspirational. Like the OpenClaw nonsense is 99.99% worthless filler where people chased a productivity hack that in reality is just hobbyist silliness.
It's token shredding for almost no value as people show that they're with it. People gloating about their swarms of agents doing effectively nothing are another "I need to max out everything" people. These are the ones who yield the result that AI has no benefit to productivity, as they overdo something to such ridiculous extremes.
The same for the laughably poorly thought out MCP servers that flood a service with a quarter million tokens for negligible value. So much insanely poorly considered nonsense is in use, to the great glee of the AI companies. And, I mean, I guess I should thank these people for basically subsidizing it for the rest of us.
The rest of us are surgically applying AI precisely, to incredible effect. The cheap plans are ridiculously valuable.