Nyt aletaan huomata, että on älyvapaata tarjota ilmaisia tai halpoja tekoälypalveluja. Alla otteita keskustelusta.
Lähde: keskustelu artikkelista
https://arstechnica.com/ai/2026/04/github-will-start-charging-copilot-users-based-on-their-actual-ai-usage/?comments-page=1#commentsI can't imagine that cliff is very far off if deep pocket entities like Microsoft are already reaching for the brakes to reign in ballooning costs. It answers why Anthropic was so desperate for "investment" in the past week or so
Yeah, I think it's a reasonably good zeroth-order assumption that under the hood, all of these things are mostly interchangeable from an internal accounting perspective. The technology is the ~same and the hardware is the ~same. It's unlikely to be the case that one supergiant organization is bleeding heavily and the other is doing gangbusters profits.
-----
Remember when people argued 'AI is only a money pit now because of all the training, but it'll be profitable on inference?'
Turns out it's the opposite. AI has a massive problem when it comes to compute use, especially with more bloated models like Claude.
It's to the point where in some cases the AI is costing more than the workers they were intended to replace.
Cost of compute is becoming an existential problem, and AI companies are getting stuck in a tough spot where their prices are too subsidized to be profitable while also being so expensive that people are already starting to look at alternatives.
------
Are these tools actually good enough to justify the costs once they are no longer being subsidized by investor money? It sounds like this could also be just the first bump as this seems to just be looking to cover the inferencing costs of just running the model and not necessarily covering the training costs. Those can be massive but could be pretty low if spread across enough usage but they also seem to be continually chasing better models which means any given model doesn't get used for as much inferencing before being replaced.
I also think the potential opacity and confusion around billing will drive some people/business away. If Joe coder can suddenly hit you with a bill for hundreds or thousands of dollars that may push some businesses away from the idea. Most places I've been have had tight controls on who can spend company money and how much. Throwing tools to people that let them create bills for $$$ without oversite seems contradictory to that. Maybe GitHub has some settings to manage that with caps or what ever else but still a consideration. Even $100 per person times hundreds or thousands of devs adds up to real money pretty quickly.
-------
Casual User (Personal Use): 5,000 – 15,000 tokens per day
Asking 3–5 brief questions, requesting a summary of a short text, or drafting an email.
Moderate User (Office / Administration): 20,000 – 100,000 tokens per day
Equivalent to: Reviewing and summarising reports, engaging in multiple extended conversations where context (history) accumulates and generating drafts for articles or documents.
Power User (without coding): 100,000 – 500,000+ tokens per day
Equivalent to: Uploading large PDF files (100+ pages) into models and querying them. This consumes large amounts of tokens because the entire document must be re-read for every follow-up question.
You have this snowball effect where the system has to read the entire conversation for every new question or request. If you have a chat with 20 exchanges back and fort, you can easily consume 10,000 tokens.
------
We’ve been warning that the other shoe would drop soon, regarding pricing, but no one listened. And now we’re starting to see their pricing models shift to funny money tokens, which have the added “benefit” or obscuring costs and making them incredibly hard to futurecast budgets for. I’m sure that part of this tokenization plan is to do exactly that, abstract the costs in a way that makes it difficult to mentally understand usager rates vs token costs.
This is especially true since it's not 1 prompt = 1 token, you have no way to know beforehand or control how many "tokens" are going to be consumed to process your prompt/request.
It's a complete black box, and if the token becomes the billable unit then the companies begin to lose the incentive to reduce token consumption and make the models more efficient because that would mean users need fewer tokens and then will spend less.
------
I got a feeling this is going to put a damper on AI usage way more than just the economics of cost and utility work out, simply because users are still charged for when the AI gets it wrong.
All those hallucinations and AI-generated errors are going to be incredibly frustrating when you pay for wrong answers and have to pay extra for the AI to try to correct them. Humans (when they're not deceiving you) are at least making progress towards working code when they get things wrong, and usually have a good upper bound of how much more time and money it'll cost to see it through. You won't get that with AI.
---------
Anyone remember the Asimov story "The Feeling of Power"? Someday, someone will discover that people can learn to do the tasks that we became dependent on computers to do for us, and it will revolutionize the world.