The era of heavily subsidised models and flat subscription pricing is coming to an end. Cursor told its users this week that Auto pricing will be charged based on the model each request gets routed to
Strong disagree. People do not want to be nickel and dimed for each and every request they send. They will actually over-pay for their services in return for a flat rate. Call it subsidized, call it what you like, the market (other than extremely wealthy businesses willing to invest the OpEx) will not pay for per-token rates.
They said the same thing about the internet 20 years ago, and everyone has essentially unlimited data plans on their smartphones, and actually unlimited cable or fiber, at flat rates. Almost no one buys a metered data plan anymore.
What will happen is that AI research will slow down, investment in new models will slow down, and existing state of the art models will get cheaper and faster. At some point we are going to plateau at the level of gpt-5.6-sol or maybe gpt-6 or gpt-7, and progress on models that are much better than that will come in years rather than weeks or months. Instead, they will focus on making these models more economical to run.
Once they can actually afford (due to hardware improvements, code optimizations, etc.) to charge a "metered" rate around $0.1 per Mtok for something like gpt-5.6-sol, a $200/month plan like ChatGPT Pro would actually be profitable for most users.
It's Kilo and your per-token business model that is written on the next door the reaper is coming for. Not the so-called subsidized plans.
I partially agree, mostly on the long term. Flat rates win once marginal cost drops enough. I think the internet story applicable here.
I do not agree on the 3-5 years date range. Coders are a finite population, and one next is knowledge workers, so adoption still has a long way to run. Every new point of adoption add token demand.
And the demand hits constraints upstream: land for data centres, rare earth production, energy generation and energy transport. I believe that extra cost will be passed to the end users of the models in some way.
BTW, Kilo's business model is to give a tool that doesn't lock in users to a single ecosystem, and not to sell tokens. My point was that everyone will pay per token sooner or later (maybe except individuals with low usage).
Strong disagree. People do not want to be nickel and dimed for each and every request they send. They will actually over-pay for their services in return for a flat rate. Call it subsidized, call it what you like, the market (other than extremely wealthy businesses willing to invest the OpEx) will not pay for per-token rates.
They said the same thing about the internet 20 years ago, and everyone has essentially unlimited data plans on their smartphones, and actually unlimited cable or fiber, at flat rates. Almost no one buys a metered data plan anymore.
What will happen is that AI research will slow down, investment in new models will slow down, and existing state of the art models will get cheaper and faster. At some point we are going to plateau at the level of gpt-5.6-sol or maybe gpt-6 or gpt-7, and progress on models that are much better than that will come in years rather than weeks or months. Instead, they will focus on making these models more economical to run.
Once they can actually afford (due to hardware improvements, code optimizations, etc.) to charge a "metered" rate around $0.1 per Mtok for something like gpt-5.6-sol, a $200/month plan like ChatGPT Pro would actually be profitable for most users.
It's Kilo and your per-token business model that is written on the next door the reaper is coming for. Not the so-called subsidized plans.
Thanks for your comment!
I partially agree, mostly on the long term. Flat rates win once marginal cost drops enough. I think the internet story applicable here.
I do not agree on the 3-5 years date range. Coders are a finite population, and one next is knowledge workers, so adoption still has a long way to run. Every new point of adoption add token demand.
And the demand hits constraints upstream: land for data centres, rare earth production, energy generation and energy transport. I believe that extra cost will be passed to the end users of the models in some way.
BTW, Kilo's business model is to give a tool that doesn't lock in users to a single ecosystem, and not to sell tokens. My point was that everyone will pay per token sooner or later (maybe except individuals with low usage).