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!
Not if they have the brain disease which makes this kind of thing appealing:
https://caviar.global/catalog/custom-iphone/iphone-17/?sort=...
Yes that flagship model incorporates an actual Rolex Daytona in solid gold.
I wouldn't go as far to call it "brain disease" though: in a sense, it is OK for someone well off to spend on expensive products (made by less rich), so things would equalize at least a bit.
Just like we in IT might happily pay 3% of our salary on slightly better shoes, and someone else would claim we have a "brain disease" because you can get perfectly good shoes for 5x less money.
> https://caviar.global/catalog/custom-iphone/iphone-17/london...
This is something that is often cited as trueism, but there is no natural laws which makes this a necessary true. There is plenty of room for black swans in market laws. So much so that the term black swan is probably better known in the field of economy then any other field.
Competition may drive down the price of LLMs, however there is a greater then zero probability that it won‘t, and if it won’t, your whole counter-argument falls apart.
But what powers the chips?
You're talking about chip economics. Inference economics requires electricity dynamics.
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.
Although there is an underlying truth: using LLMs for large-context tasks like coding is still extremely expensive.
Didn’t happen for me.
On the Plus plan newer models reached the limit faster so less tasks where done until I had to wait 5 hours
"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’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.
My guess is that AI will be more like consumer electronics than like Uber.
You can get a table from Ikea that costs a fraction of what an artisan makes. They're not the same final product but their functions is the same.
Anyway, it might be worth it to invest in an LLM rig today if you’re paranoid.
But I think what matters is that the new generation of coders will adopt it as the norm. Gone are the days where you download a free text editor and just trial and error with the documentation one tab away. Every bootcamp is teaching react with clause and cursor. You have to pay to for a subscription to build your BMI calculator.
[0]: https://idiallo.com/blog/paying-for-my-8-years-old-ride
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.
Yes, but every high performing open weights model coming out of China has (supposedly) been caught distilling frontier models.
It seems like a lot of people are making assumptions about the state of the open weights ecosystem based on information that may not be accurate. And if the big labs are able to reliably block distillation, we could see divergence between the two groups in terms of performance.
The big labs will not be able to reliably block distillation without further inhibiting general use of the models, which itself will help tip the balance away from commercial models.
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.
In my experience, to safely get any value out of an LLM, you have to be more knowledgeable than the LLM on a topic. So in this case, you'd really need a PhD to use this tool, so at best its a $20k a month research aid, which honestly is far more expensive than a handful of grad students, and probably less effective.
Whether they can deliver is another question, but I wouldn’t bet my career that they can’t.
> 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.
Sounds like a lot, but in the few weeks I've had it, I was able to complete two projects I had given up on due to not having time in the past. I re-jiggered some other monthly subscriptions so the net cost wasn't ultimately that much more than what I was paying previously. I also weighed it against buying something like a DGX Spark for local inference, but ultimately I don't want to mess with serving models (and the ones available just aren't as good, realistically), I just want a good one that works.
I probably can't justify much more than $200/mo, but for what I get out of it, I'm happy to pay it. I've done more in the past few weeks on side projects than I had in a couple years.
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...
So maybe it's true you won't get "the best", but I don't think you'll be that far off.
The "good enough" points so far have been
- "as good as ChatGPT"
- "as good as GPT4"
- "as good as Sonnet 3.5"
- "as good as Opus 4.5 or Codex 5.2"
Anyway, we'll see where the chinese models are in a year, and we'll see where my expectations are. Hopefully they overlap at some point.
every time you see "open models" read: "binary blobs impossible to build locally temporarily not monetized"
If there is an AI boom, what we're seeing is its infancy. Semi-autonomous coding is the first and most natural use case, thanks to vast amount of training material, opportunities for closed-loop RL with minimal human supervision and the eagerness of the community to try and embrace new tools.
But it is still not much more than QoL improvement at this stage, and maybe some velocity gains for the most hardcore users willing to spend time and money to stay at the bleeding edge.
But there is also a rather large appetite for local models, I am not sure the future of AI will be 100% cloud based.
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.
I wrote more about this in a blog, at https://www.viblo.se/posts/ai-hobbycoding/
But this is absurd:
> the cheapest usable tier of Claude Code is $100/mo
If you pay by the token instead of with a subscription, and don't send the entirety of your code base with each request, costs are ridiculously low. Like, $50 will last a minimum of 3 months of heavy use on openrouter.
It's also far from certain everyone needs the latest version of the best "frontier model"; it very much depends on what you do.
However, on a fixed price plan your behavior changes. It's a qualitative change in how you work, rather than quantitative. Ideation and product design and specification start becoming bottlenecks.
I started out the API route. I started spending $100 a month once I was spending upawards of $10 in tokens a session.
But sure, lets just keep automating ourselves out of jobs (and help other industries do it too) with no plan as to how to help all the displaced people.
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.
I mean, you have to declare when content is an advert, and if you are asserting that the owners of the chatbot are going to just ignore that requirement, won't they just do the same thing for paid accounts?
Unless you’re a top-tier domain expert. Then you’re safe until (if…) ASI.
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.
Be those people.
Businesses will pay more since they can justify the cost. That seems fine?
There was no such time. Even if everyone means "every software engineer" or any variation thereof, and we substitute any other such tool for GC.
"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.
"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?
The author doesn't even mention Codex even though it likely will out compete Claude Code.
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.
Just think how these big companies will use that kind of power for themselves to get even more extreme uses out of it.
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
