DEFINITELY NOT A BUBBLE EVERYBODY NOTHING TO SEE HERE
assuming API pricing
API pricing is not what it costs openai to offer the inference.
API pricing is what they’re currently selling their inference, with profit included in price.
“But if the user didn’t get a subscription, he would have spent $14k in token” is a wrong assumption.
The user is wasting the tokens on useless tasks because they’re free to him, but once he has to pay them, he wouldn’t ask to generate 100 images of a toilet eating carrots anymore
Headline is innacuralte. Customers who purchase a $200/month subscription, can save up to $13800 on token costs compared to “a la carte” token pricing. Open AI does lose over $1 per $ in revenue, but this is more a function of them having too much compute, and very high training costs, rather than a loss per gpu hour on tokens served, it should stilll be a loss for OpenAI:
Open source model pricing per gpu hour, ranges from 80% margins at high batch saturation per user (low tps per user) to 50% at concurency of 8, to loss at fewer user requests per gpu. https://inferencex.semianalysis.com/compare/deepseek-r1-b300-vs-h200
OpenAI has very high prices, and too much compute. They would likely lose $6000 per user who maxes out their $200 plan, and certainly over $1000 per user.
Is this a reverse psychology trick to convince people to pay for ChatGPT subscriptions?
If you have it through work you know what to do
Welp, just got fired for sending 18000 prompts to ChatGPT today.
Rookie mistake. Send 18000 Codex prompts next time and they’ll hail you as a god amongst men
I salute your sacrifice.
Totally not a bubble, honest guv!
I’d rather save earth
I dunno, man. I’ve blown $200 on worse things.
Then don’t use the internet because everything has Ai. Google has it, every browser has it installed, almost every shopping site automatically uses Ai results, the news feed on Google is all Ai results, the questions that are the most asked on Google are now Ai driven.
Main point of environmental destruction is using the LLM’s but nearly everything is using Ai and it’s hard to outrun it and a lot of things are making it hard to opt out and I wouldn’t put it past them to make it so they can make you pay to opt out of Ai in the future for a premium browser with the ability to disable Ai otherwise you’re stuck with it kind of thing.
I hate what it’s doing to the environment too but it’s not going away unless everyone in neighboring communities decided to bulldoze them and use their wrecking balls to destroy them. Emps aren’t hard to make happen.
If people wanted to destroy them they would. If people really didn’t want them they’d destroy them.
every browser has it installed
Not Waterfox!
What bubble have you been living in? Almost none of my apps have any kind of AI… You can easily ditch it by using a Firefox fork, or any privacy Chromium fork, or even Chromium itself. Saying “Then don’t use the internet” is somewhat of a defeatist sentence, just stick to good websites/programs and you’re good.
I should’ve added a 🤪 to that because that was meant to be light not in a fuck off kind of way but more of a you gotta be kidding me this shit keeps expanding kind of way.
Oh
Use Firefox without AI and don’t use google search engine
How much of our water, electricity, tax breaks, and public land does it use?
Yes!
Nice try, OpenAI sales reps.
But how much is the data you’re giving them worth? The other option is don’t give them your money or your data. The Qwen 3.6 MoE model with OpenCode is running pretty well on my RTX 4060 gaming laptop. According the Codacus YouTube channel, it even runs decently in as little as 6GB of VRAM.
Edit:
Fixed typo.
TBH local models aren’t as good as cloud. Even with 16GB VRAM you aren’t getting anywhere close to >100GB cloud LLM
No, it’s not quite as strong, and especially the initial prefill can take a bit. I also sometimes run into infinite thinking loops where I have to stop it and re-run my last prompt.
It’s surprising how close Qwen 3.6 gets on the benchmarks to Claude models, though. Especially when running locally with 200k context, I’ve found it’s good enough to be a daily driver. Despite the faults, it’s better than paying Anthropic $200 a month so they can rate limit me and collect my data.
I prefer to run with cheap pay-per-prompt cloud model. You can find really good open models that cost $0.50 per million tokens.
Not $14k lmao
No, but do you really want them collecting your data? Your prompts and anything else you share is helping them quite a lot.
That wasn’t the question that was asked.
Your data probably isn’t worth $14k to them in most cases, no. If you’re using it for legitimate purposes, though, still helping them by using their product and sharing your data with them. They’re getting data showing how a real human thinks and interacts with their system, and if you have it help you with any of your own IP that isn’t AI generated, you’re giving them something that is literally priceless.
“Literally priceless” that nobody would be willing to pay $14k for.
I have code for personal projects that solves problems in novel ways as well as other creative work that I don’t care to let Anthropic and OpenAI train their models on. Is my work worth $14k to them? Well, the value is intangible to me, and I can say at least that companies have paid me a lot more than that for code that took a similar amount of time to write. The major data sources for training LLMs like GitHub, Reddit, Wikipedia, etc. have already been tapped, but they always need more and more data. If you want to give them your data like it’s not worth anything, you do you, but they’re not getting mine. If I need LLMs for personal use, it’s local or nothing.
You keep trying to make this into something it’s not. All I have been saying is that the data is not worth $14k. You seem to want to make it about me being okay with them having access to my data. I never said that.
Just stop using it period, self hosted or not. Wtf is the thinking here?
It’s 100% bad in every case. It’s never good, is never been good, and it literally can’t become good. It’s bad no matter where it’s at, or who it’s being hosted by.
Just don’t use it for anything. At all. Nope, not even that.
I started out using GitHub Copilot at work because there was a lot of pressure to use AI, and I was put off by how we were churning through PRs that seemed to work, but having to go back and fix the slop afterwards.
Now I’ve realized that there are skillful ways and unskillful ways to use LLMs, and they can in fact be a useful tool beyond just generating slop. They don’t replace a human thinking critically, but they can automate mundane, routine tasks. They can also summarize text well and suggest options for humans to consider. For example, LLMs reviewing code will often find issues the human reviewers missed.
In addition to coding, I’ve recently been using Qwen locally for screenwriting. It can’t write worth a shit, but it does a good job critiquing my work and pointing out problems with the story structure and the like. For example, I can tell it something like “look at the 7 plot elements described in this MD file and point out where this story does and doesn’t follow this structure”, and the output is quite useful.
While LLMs aren’t the magical silver bullet the tech bros are hyping them up to be, they can still be a useful tool. If they’re just used to generate slop, then no, they’re worse than useless.
I’m horrrfied that an LLM is your writing coach.
It’s not my “coach” any more than random people online would be if I posted it in a forum somewhere and no more than a LLM or a human peer reviewing my code is my “coach”. It provides a different perspective to help me see beyond my own biases with feedback I can accept or reject.
Qwen has obviously been trained on writing books and a ton of screenplays. As an experiment, I changed the character names in a classic sitcom script and it was able to identify the series from the writing style and then it also identified the episode. It’s not useful for doing the actual writing, but it does provide useful feedback based on sophisticated statistical analysis of my work compared to its professionally-written training data.
Explain to me how it’s better than you learning to analyze your own work from a formulaic perspective?
Everytime you choose to use AI, you are choosing NOT to develop an ability of your own. Sometimes, that’s an ability that just tedious to use, other times it might be something you obviously need to do yourself, yet others the ability might be something with a tangential utility you haven’t recognized.
An analogy might be reading music exclusively. Great, now you can play a wide range of music–indisputably beneficial!–but the cost of developing your own ear.
I have read a lot of books and do analyze my work in terms of techniques and principles I’ve studied over the years. However, even top professional writers don’t work in a vacuum. TV writers, for example, have “the room” with a team of professional writers, producers, etc. weighing in on all writing decisions. For indies, you don’t have that luxury, and even getting another human who is good at writing to read what you wrote and share detailed feedback is hard, especially when said humans aren’t getting paid to do it full time. Asking friends and family to critique your writing will often result in them trying to spare your feelings, whereas Qwen will happily rip your work to shreds and not care if it just shit all over your passion project.
Do you have a flat rate sub to Qwen? I’m curious if you fed it something that you personally think is great writing that isn’t prominent training data, that you are intimately familiar with, and what you would make of its analysis?
My fear is two-fold: first, writing is communication between people with shared experiences. An LLM can’t really tell if someone’s going to have an emotional connection to your writing or why or what or how it works. Second, novelty and rule-breaking is highly context dependant. I’d be worried an LLM is merely steering me into probable lanes instead of allowing me to develop my own unique voice.
Well, I guess I’ll buy that subscription then
Just checked Api pricing: it’s 30USD for million output tokens. Heavy usage will absolutely go into hundreds of dollars of costs, input tokens are $5/1mil so if you input files and long context it’s going to get even more expensive, potentially upwards of $1000 per month.
Please do.
Use it all.
Bankrupt these shit companies and help burst the bubble
This is their strategy, they want people to use it, get hooked, replace parts of their day-to-day life with it, make it to difficult to “just go back”, then hit them with the actual bill.
They won’t go bankrupt unless their backers walk, and their backers are still quite confident in this strategy… because it’s working.
I’m thinking of getting a subscription and burning tokens out of spite.
But no one is going to be able to afford a $14,000 subscription for slop.
No, but if you get that number in everyones heads, there is much less resistance to, say a 100 EUR a month increase.
Yes, but what good is that if 100 USD more is still losing tons of money.
“A $200 ChatGPT Pro 20x subscription could cost as much as $14,000 in API pricing if fully utilized.”
What is being said is that in a month time you can burn through $14,000 worth of tokens. It does not show that that comsumption would cost ChatGPT the full $14,000.
So what they are doing here is planting the idea that a $200 subscription is actually worth $14,000. Which makes if very easy for them to make people switch to tokens or just increase the price of your subscription without losing too many customers.
It is generally thought that official token price is breaking even, at best. Maybe it is not even doing that. The worth of the AI usage has nothing to do with those costs and is likely much lower. So far there is very little done on understanding actual value of AI use though.
I am not saying you are wrong, just that the above statement can be correct at the same time.
companies can
what does a human employee cost per month?
Sounds like a trap. Big cruises are said to have buffets, but yet, they’re still floating.
Gov will bail them out and use our taxes to do it
I wonder what companies that have integrated AI into all their workflows and processes are planning to do when the times comes to pay real price for the tokens.
spoiler
Nothing. They aren’t thinking ahead.
Nah, they do what they always do: mass layoffs
Lindy, announced that the company moved 100% of its traffic to DeepSeek V4, switching entirely away from Anthropic’s models. DeepSeek V4 proved comparable to Claude Sonnet at a fraction of the cost
They move to a cheaper and shittier AI. The answer is unfortunately not that they re-evaluate human workers and create good employment opportunities
That’s the next CEOs problem to solve while the current one is enjoying his golden parachute and sailing around the world. Right now, number is going up!
The companies don’t pay the price, they just pass it on to the consumer with a markup. Right now they just try stuff out to see what people really use AI for. Eventually the “AI features” will be cut back to the parts that really make them money, once they have to pay the real price.
All the investors know it’s a massive money sink right now. The goal isn’t for “everyone” to get to use AI.
It’s to get so many people used to using AI that businesses like law offices and hospitals and other corporations so ingrained and built around having AI, while leaving so many graduating college students useless without AI, that businesses will be reliant upon it, no matter what costs of it they will have to absorb.
In five years there won’t be a $200 plan. There will be a $15,000 plan per person and businesses will pay it because they won’t be able to do well without it.
Their investors better hope that’s not their plan. It’s a terrible business plan, local open AIs will crush any attempt at profit. Local open AIs would already be crushing profits if business were run sensibly. You don’t have to train a whole model from scratch. You can just give a pre-trained model your toolset and prompts (things you’d’ve had to make anyway) and call it a day.
Law offices, hospitals, and insurance companies are not going to set up their own servers, localize an AI, and upload all the custom data sets they need into a pre trained locally hosted model. Everyone could easily host their own cloud backup system as well, but no one does it.
But yes. That’s there plan. It’s been done multiple times before amongst things. It’s how busses eliminated trolleys, how ride share beat out taxis, and how Walmart sells the cheapest Coke. Amazon does it as well. It’s a variation of predatory pricing. It’s common as hell. Supposed to be illegal, but seldom does anyone get in trouble for it. Crooked government that works for the wealthy n all.
Sure, but they’re not hiring Open AI to do custom work. They’re hiring some third party who well might use their own servers. Even if they are hiring Open AI there will be a thousand competitors in no time. There’s no meaningful cost barrier to set up a local AI. Sure scaling costs a little, but really not much. It’s a foolish investment that only works if people are incompetent… okay so might work a little, but it’s still stupid.
“I can’t quit my job because I’ll lose my
health insuranceAI access.”I think there may also a horizontal scheme as monopolies take on a global scale. Those businesses that sell in bankruptcy due to high tech costs could be gobbled up by the biggest AI-native competition. It’s a leap but maybe in a decade your optometrist is replaced by an ai kiosk with a remote technician?
Maybe then he would give me the prescription I paid for…
If you’ve got a toy project that you want “AI” to give you a hand with, do it now.
Pretty soon all these companies are going to have to pay for all that investment in compute resources they’ve been busily soaking up over the last few years, and then they’re going to have to pay back their investors, and then they’re going to have to try and make a profit
This is the golden time for cheap commercial AI. Already the noose is starting to tighten, and it will never again be as cheap as it is now.
Yes we’ve begun to track “token use” all over my company so it doesn’t spiral out of control, as it easily can do when you have agents managing agents connecting to MCP servers that themselves use the models to generate responses. The engineers around me say that they basically have multiple agents cranking full time and just keep an eye on them every so often. They will even queue up things to run overnight to make use of the time. They never actually close their laptops. This is an insane amount of usage, well beyond what anyone can do in the ChatGPT application by typing with their fingers, and there’s no way it can continue like this.
Unless there are actual major efficiency innovations. But yes, current LLMs are sold cheaper than what they cost
Sounds like it’ll never be worth it.
In five years once this RAM nonsense is over you’ll be able to run a comparatively high quality local LLM for very little money. I can’t see how these companies will ever make their money back.
If manufacturers are willing to sell components to us in five years that is.
Of course if the colllapse happens before then the story might be different…
I’m slightly optimistic that manufacturers will return to the retail market eventually. Every AI company is racing to hyperscale right now but there will be a point where the infrastructure is built and at that point the growth will slow down quite a bit. In that scenario there will be ongoing demand for components to be replaced as they become obsolete but I can’t imagine the demand will be the same level it is right now as everyone rushes to build.
That’s assuming this all works the way they want it to. If the economics aren’t viable and the bubble bursts…
“Hyperscale” is utterly meaningless MBA jargon at this point. Equivalent of verbal slop from industry shills and CNBC/Bloomberg sell side simps.
Sorry if that’s true. I understood the word to mean aggressive growth at any cost to try and shut out competition before they can get established.
Their Datacenter buildout doesn’t work they want to. Most projects are very much delayed, and those that even started getting built are over budget. OpenAI and Anthropic will collapse in the next years, and this is coming from someone who absolutely sees the good things about the technology itself.
OpenAI and Anthropic will collapse in the next years
Stop, I can only handle so much good news!
There is no way, absolutely NO WAY to recuperate the amount of cash burnt on those two companies, and that is not even counting the amount of AI Startup whose cash is currently flowing towards to those two.
🤞
Sounds like price hikes to communicate costs are coming and resources are going to be redistributed to productive uses.
deleted by creator
This is the golden time for cheap commercial AI.
I suppose, but small open weight models with more advanced coding frameworks optimized for them are catching up fast and you can do it privately at home on a mostly affordable consumer graphics card.
If you have solar it’s basically free, minus the graphics card CapEx you may want for gaming anyway, as well as some setup time and a bit of patience.
Yes, it’s trending in that direction, and I’ve been experimenting with pretty small models on my PC as I don’t really have the hardware to go large. If you’ve got the coding chops to set it up, it’s definitely something to keep an eye on.
There’s actually scope for someone to set up / sell local compute hardware+software packages, similar to all those coin miners. Give the end user a way to update models, or push models out to them or something, it seems it would be a good middle ground between manually typing code like a peasant and total corporate AI apocalypse.
There’s actually scope for someone to set up / sell local compute hardware+software packages, similar to all those coin miners.
I think that’ll be a viable target in the future, and have little doubt some are jumping on it already. However, I also think it’s too much of a moving target currently, a near optimal setup changes almost entirely month to month.
I find myself targeting last months setup, as then there’s enough literature out there to get it set up in a day or two and most of the kinks have been worked out. Otherwise, I lose too much coding time to debugging the bleeding edge.
IMO, at the moment, if you’re not capable of setting it up yourself you likely don’t have the experience to use it reasonably safely nor an adequate understanding of its limitations. You’ll find yourself using more time fixing the blunders than you gain, and / or the project will spiral out of control in maintainability, security, readability, and so forth. You could get away with small projects written as ‘write only’ code ala Perl though, keep the prompts and tests, when it needs to change rebuild with the newest hotness. Inefficient and unsatisfying though.
What’s your setup, if I may ask? I’m using llama.cpp router with vscode kilo.ai and qwen3.6-35B-MoE-MTP as a model mostly. It’s surprisingly good as a coding assistant, but I think you have to know what you are doing and know your stuff(aka be an experienced developer) to make it useful. just letting it vibe leads to crap code
just letting it vibe leads to crap code
Yup, vibe is occasionally useful for proof of concept stuff, but disastrous for maintainability, security, readability, or large codebases. Without experience it’s still a foot gun for anything even slightly serious.
Best approaches for a learner are to consider it autocomplete that needs research. Look up what it’s suggesting, see if it’s hallucinating, with luck it’ll point you in a useful direction where you can learn a good solution, as it has no idea what that is. Also makes a pretty good rubber duck for hashing out architectural decisions, finding alternative approaches etc, though you’ll have to point it at a web search for that. Spin up an e.g. vane instance for this, as small models don’t have enough world knowledge. Use it to write (or preferably copy from its system prompt examples) boilerplate and unit tests, perhaps descriptive comments (doublecheck).
One thing to do is put everything you learn about coding style into your system prompt as they’re dogshit at consistent style without significant beatings around the head. Finding your own comfortable, consistent style is super useful for future readability. The joke about when I wrote this only God and I understood it, now only God does, will come clear in a month or two. Learn to work around it. Simple beats fancy unless you truly need the speed.
While I do use agent iterative approaches, probably best to approach that organically as you grow, monsters lurk there. If you must, containerize / vm / isolate the hell out of something like opencode to muck around with.
FWIW I still write most of my code by hand, it’s simpler and more consistent, but I’m keeping an eye on the development of LLMs, and I will let it write scut code (that I edit later). Code and Mathematics are super structured languages, pretty much ideal for large language models, so I can see them maybe, eventually getting good. More general thought, not so much without significant architectural upgrades.
While this advice is true for all models, when it comes to agentic tasks (add this small feature/write this test harness/find bugs/suggest improvements), open source models are still way behind, vibe code or not.
Claude Fable or even Opus in an editor like Zed have a 1 million token context window and will “think” through the goals of the application, test their changes, work through debugging processes the way a programmer would, stop to ask for clarification, check diagnostic tools and linters, prompt to run test code, etc.
Llama, Gemma and Qwen etc. Do lack a lot of the world knowledge to get the goals of the application, but they also just don’t have the debugging skills, won’t test their code, don’t always tool call correctly, get confused as the context increases and nobody has enough vram to run on large context sizes locally.
They can do autocomplete on small functions but aren’t really there for more complex tasks.
On top of that, the biggest problem is that the best open source models are trained and released by the same giant tech conglomerates that have an interest in not competing with their own products. Qwen is Alibaba, Llama is Meta, gpt-oss is OpenAI. Even the more “independent” ones, kimi (Moonshot) and GLM (z.ai) are mostly funded by Alibaba and Tencent. They’re released for research and marketing purposes and to please their corporate backers with inflated stock. Almost nobody has the resources to train new models from scratch. People make lots of merges and fine tunes but AI is not democratised the way that traditional programming tools have been.
Maybe some day there will be enough cheap compute for open source communities to pool together resources to build competing models but they’re not really there yet :(
Context management is a huge part of making smaller models viable (and likely a big part of making frontier models better). Tricks like structured context libraries for thinking improve things a lot, I like approaches that output things like an Obsidian vault that let you dig in and correct bad assumptions easily, even if it’s a bit slower. It’s a useful deliverable that can (mostly) be reused with updated models.
Things like ‘the debugging skills, won’t test their code, don’t always tool call correctly’ are tangibly improving model to model, framework to framework, and are problems that will be solved in time, but yes they need handholding ATM.
Things like
test their changes, work through debugging processes the way a programmer would, stop to ask for clarification, check diagnostic tools and linters, prompt to run test code
are mostly down to framework, not model (except for failing to tool call, which is improving), and falling at a respectable rate.
That said, sure, frontier models get more in one go, personally I’m fine with only a 3-4x force multiplier instead of 10 to keep it local, but YMMV. For a business with resources for a bigger server it’ll be more like 8 times. Remember that some businesses handle sensitive data and can’t (or damn well shouldn’t) use frontier models, so the market is there.
Maybe some day there will be enough cheap compute for open source communities to pool together resources to build competing models but they’re not really there yet :(
Not wrong, decentralized inference is mostly solved (with latency penalties), but without decentralized training true democratization will remain out of reach. Hopefully a breakthrough will ensue, but until then we are dependent on the kindness of corporations (or them rugpulling competitors).
This could also be a part of the RAMpocalypse thing, ‘if there’s not a moat I’ll fucking dig one, damn everyone else’ (and damn SamA). I doubt that’s sustainable long term, but it might get them through to IPO, more’s the pity.
Yeah they’ve been pushing Claude code at work for us non coders jobs to come up with stuff that would help us. We’ve gotten a few surprisingly useful programs out of it, but our assumption is perfect them now before pricing goes through the roof. We are also only creating programs that do not require ongoing AI use. Just a bunch of relatively simple things that make our jobs easier.
I am still pushing my boss for some local hw as I think as a group we’ve spent a couple grand in the last month and that is the least of my reasons for wanting a local llm vs subscription.
Another way to look at it would be “if you’ve got a toy project to practice coding without AI on, do it now” before that is the only option.
This is just Gym Economics though, right? They work on the assumption that only a small number of their member will actually use the service heavily, but the overwhelming majority will turn up to use the treadmill a few times then never visit again.
Ok but it would take 70 users paying $200 to cover the cost of $14,000. So if one person maxes out their usage, there needs to be 69 users who do not use their account at all but are still paying. And that’s just the break even point, still no profit for the AI company.
I’m struggling to believe that many people would pay that much and then underuse the subscription. It seems far more likely to me that this pricing model isn’t sustainable.
For a consumer service, absolutely. That’s too much money for a household to ignore and they are actually paying attention.
For a business user? Quite possible. My company bought subscription to one of the providers for every single employee, no matter the role. A large number don’t use it at all (if they do anything, it’s using a chat that’s either free or included with something else), and most of the rest use it lightly. We have only a handful of folks trying to use it as much as possible. Companies frequently just buy for everyone instead of micromanaging who needs or doesn’t need a service.
I thought that Anthropic et al. were charging enterprise accounts based on token usage rather than just a flat subscription fee. That’s why you see things like this.
Even worse, that calculation is based on that their API pricing is currently providing a positive margin. From what I have seen and heard at this point, API pricing is at best breaking even.
