One thing I rarely see discussed is that AI cost is not just dollars per token.
There’s also latency, dependency on external infrastructure, privacy and compliance concerns, energy usage, and just the general predictability of the system itself.
My guess is that this will gradually push a lot of companies toward more hybrid architectures over time. Small or local models are probably good enough for things like filtering, routing or repetitive high volume tasks, while frontier models get reserved for the places where the quality jump actually justifies the added cost and complexity.
As useful as frontier models are, using them for absolutely everything sometimes reminds me of using a distributed system for problems that could have been solved locally with something much simpler.
I wouldn’t be surprised if, in many real world cases, a fast specialized system plus a smaller model ends up being the more practical and economical setup overall.
What always happens. A market correction followed by going back to a reasonable state, until the next bubble of course.
In my opinion, LLMs are useful for many things but not anything and everything and definitely not in the way the boosters are claiming. This is not a popular opinion when you are inside the bubble or have something to gain by it. So when there there's a downturn, things will hopefully stabilize with LLMs being another tool that can be used to automate certain things. It feels crazy saying this these days and have been told I'm out of touch if I think this way and who knows, maybe that's true.
Less people will use the frontline models and those who do will pay more. Progress will slow. OpenAI will sell your chat data. You will get an AI tax. Companies will use less of it.
Hopefully new ways to deliver similiar quality will be discovered.
Same pattern as cloud pricing free, then efficiency pressure kicks in. When costs rise people will move from heavy LLM usage to smaller, task-specific AI and more optimization.
Even if one provider raises prices to the point where things become unsustainable, alternatives tend to emerge eventually. Chinese LLMs, or maybe someone else.
Personally, I'm hoping the next breakthrough won't just be another Transformer-based LLM, but a fundamentally more computationally efficient architecture.
For a lot of companies, probably shut down or drastically limit their AI usage due to rising costs. A small or medium sized business dependent on ever growing AI expenses is in a real bad position, and could well go under.
I heard a few companies ended up going back to hiring actual employees for work that was previous done by LLMs, so there's a chance we could see some more of that too. Might also see a few try to make it work with outdated or local ones too.
I wonder though what is the "good enough" level of LLM assist for software dev?
Similar to say CD quality audio or 4K video is good enough for most people.
I feel that Claude Sonnet is nearly "good enough" for my current workflow, at least while I am still in the loop, interacting and reviewing code manually before committing.
I think we might finally move away from subscriptions in software. You don't expect your toaster company to pay your electricity bill for you, but we do that for our software apps. I think that the rising token costs will adjust the way we consume software. Consumer behavior will shift to pay for AI as utility and software apps will compete to be more token efficient.
Token anxiety is real. What worked for me:
prompt caching on fixed system prompts cut my Anthropic
bill by ~60% overnight. Most devs don't realize cache
writes are 25x cheaper than input tokens on Claude.
Local models for classification/routing + frontier only
for generation is the other move — but the latency
tradeoff is real if you're in a user-facing flow.
Sometimes I do wonder about this. Some companies might get people used to AI first and then raise prices later, which could put many of us in a difficult position. But I also think Linux came out in a similar kind of environment, and in the end the community will find a way through it.
It’s not just cost per seat. It’s lock in, eroding skills, latencies. I worry about this a lot. There are companies that rely on Claude or Cursor in a way that is not easy to rip out, even if rates 10x.
Frontier price will keep going up as AI gets smarter and can be applied for more economically useful tasks. That isn't the same as "AI pricing is going up" because intelligence per dollar has consistently cratered and will continue to do so. You just won't use frontier intelligence just like you don't use industrial equipment in your house.
What do you think will happen? How does supply and demand work? Practically every business and government in existence is existentially dependent on AI, speculation on it is the only thing keeping the world from global financial collapse. It's "too big to fail" at a scale that dwarfs the financial crisis of 2008.
You'll pay the fucking danegeld is what you'll do, and keep paying it, because you reorganized your entire existence around and mortgaged your future on a closed proprietary third party service's business model that is now a single point of failure for our entire technological civilization, making its market value practically infinite.
That's a collective "you" there, by the way, not "you" personally.
Isn't it strange? You'd think there were some lessons learned from the 2008 crisis but apparently not. It is not that long ago to be forgotten already.
The lesson is that if you’re too big to fail no laws apply to you and there unlimited money to be made.
It has been learned very well.
The brazen violation of intellectual property was a precondition of making this technology useful. Taking the risk of breaking the law at this unprecedented scale was an informed decision made based on this very lesson.
One thing I rarely see discussed is that AI cost is not just dollars per token.
There’s also latency, dependency on external infrastructure, privacy and compliance concerns, energy usage, and just the general predictability of the system itself.
My guess is that this will gradually push a lot of companies toward more hybrid architectures over time. Small or local models are probably good enough for things like filtering, routing or repetitive high volume tasks, while frontier models get reserved for the places where the quality jump actually justifies the added cost and complexity.
As useful as frontier models are, using them for absolutely everything sometimes reminds me of using a distributed system for problems that could have been solved locally with something much simpler.
I wouldn’t be surprised if, in many real world cases, a fast specialized system plus a smaller model ends up being the more practical and economical setup overall.
[dead]
What always happens. A market correction followed by going back to a reasonable state, until the next bubble of course.
In my opinion, LLMs are useful for many things but not anything and everything and definitely not in the way the boosters are claiming. This is not a popular opinion when you are inside the bubble or have something to gain by it. So when there there's a downturn, things will hopefully stabilize with LLMs being another tool that can be used to automate certain things. It feels crazy saying this these days and have been told I'm out of touch if I think this way and who knows, maybe that's true.
Less people will use the frontline models and those who do will pay more. Progress will slow. OpenAI will sell your chat data. You will get an AI tax. Companies will use less of it.
Hopefully new ways to deliver similiar quality will be discovered.
Stock market will pop.
Prices will go up for people inside the moat
Same pattern as cloud pricing free, then efficiency pressure kicks in. When costs rise people will move from heavy LLM usage to smaller, task-specific AI and more optimization.
Even if one provider raises prices to the point where things become unsustainable, alternatives tend to emerge eventually. Chinese LLMs, or maybe someone else. Personally, I'm hoping the next breakthrough won't just be another Transformer-based LLM, but a fundamentally more computationally efficient architecture.
For a lot of companies, probably shut down or drastically limit their AI usage due to rising costs. A small or medium sized business dependent on ever growing AI expenses is in a real bad position, and could well go under.
I heard a few companies ended up going back to hiring actual employees for work that was previous done by LLMs, so there's a chance we could see some more of that too. Might also see a few try to make it work with outdated or local ones too.
I wonder though what is the "good enough" level of LLM assist for software dev? Similar to say CD quality audio or 4K video is good enough for most people. I feel that Claude Sonnet is nearly "good enough" for my current workflow, at least while I am still in the loop, interacting and reviewing code manually before committing.
I think we might finally move away from subscriptions in software. You don't expect your toaster company to pay your electricity bill for you, but we do that for our software apps. I think that the rising token costs will adjust the way we consume software. Consumer behavior will shift to pay for AI as utility and software apps will compete to be more token efficient.
Token anxiety is real. What worked for me: prompt caching on fixed system prompts cut my Anthropic bill by ~60% overnight. Most devs don't realize cache writes are 25x cheaper than input tokens on Claude.
Local models for classification/routing + frontier only for generation is the other move — but the latency tradeoff is real if you're in a user-facing flow.
How did you do that?
Prices are going down. Just look at open source models, you can run the equivalent to a SOTA model 8 months ago on your laptop.
Sometimes I do wonder about this. Some companies might get people used to AI first and then raise prices later, which could put many of us in a difficult position. But I also think Linux came out in a similar kind of environment, and in the end the community will find a way through it.
It’s not just cost per seat. It’s lock in, eroding skills, latencies. I worry about this a lot. There are companies that rely on Claude or Cursor in a way that is not easy to rip out, even if rates 10x.
I think it’s going to be like infrastructure —- eventually they will reach certain level, maybe like electricity.
Frontier price will keep going up as AI gets smarter and can be applied for more economically useful tasks. That isn't the same as "AI pricing is going up" because intelligence per dollar has consistently cratered and will continue to do so. You just won't use frontier intelligence just like you don't use industrial equipment in your house.
most people will stop paying for the frontier models and will look out for the small models which are optimised on certain tasks
What do you think will happen? How does supply and demand work? Practically every business and government in existence is existentially dependent on AI, speculation on it is the only thing keeping the world from global financial collapse. It's "too big to fail" at a scale that dwarfs the financial crisis of 2008.
You'll pay the fucking danegeld is what you'll do, and keep paying it, because you reorganized your entire existence around and mortgaged your future on a closed proprietary third party service's business model that is now a single point of failure for our entire technological civilization, making its market value practically infinite.
That's a collective "you" there, by the way, not "you" personally.
Isn't it strange? You'd think there were some lessons learned from the 2008 crisis but apparently not. It is not that long ago to be forgotten already.
The lesson is that if you’re too big to fail no laws apply to you and there unlimited money to be made.
It has been learned very well.
The brazen violation of intellectual property was a precondition of making this technology useful. Taking the risk of breaking the law at this unprecedented scale was an informed decision made based on this very lesson.
AI model prices are getting cheaper over time, per amount of capability.
opus 4.6 -> 4.7 did not respect this assumption
4.7 is better imo and uses fewer tokens per requests
and it did not even got better.
Most companies will fail
[flagged]
[flagged]
[dead]
[dead]
[dead]