GitHub Copilot Users React To New Usage-Based Pricing System (arstechnica.com) 27
An anonymous reader quotes a report from Ars Technica: In April, GitHub announced that it was moving subscribers from request-based billing to a usage-based model for its AI-powered Copilot service. As that new pricing model goes into effect today, many GitHub Copilot users are reporting some extreme sticker shock as they realize just how quickly their previous "normal" usage is burning through their newly limited monthly allotment of AI credits. Across social media and forums, many Copilot users are sharing personal statistics showing how just a few hours of AI usage can now account for a large chunk of their new monthly subscription caps. For some users, it reportedly took less than a day to use up a month's usage quota.
That's a big change from previous months, when GitHub Copilot subscribers were allocated a certain number of "requests" and "premium requests" based on their payment tier. GitHub said that the old system meant that "a quick chat question and a multi-hour autonomous coding session [could] cost the user the same amount," forcing Copilot itself to "absorb much of the escalating inference cost behind that usage." [...] Indeed, some Copilot users have been sharing estimates from GitHub's own tool showing that their previous monthly usage would rack up bills in the thousands of dollars under the new pricing plan. Under GitHub's new usage-based pricing system, paid Copilot subscriptions instead grant users a certain number of AI "credits" each month, with one credit corresponding to $0.01 of usage. Subscribers also get bonus credits depending on their subscription level: the $10/month Pro plan includes 1,500 credits ($15 worth); the $39 Pro+ plan includes 7,000 credits ($70 worth); and the $100/month Copilot Max plan includes 20,000 credits ($200 worth).
The precise number of Copilot credits used by a given prompt is determined by the number of input and output tokens used and the rates charged by the underlying large language model. That means pricing is highly dependent not just on the type of request but on the specific model that a user chooses. One million output tokens from OpenAI's GPT-5.4 nano would run just $1.25 on GitHub Copilot, but that same level of output would run $30 on the frontier GPT-5.5 model (Copilot users who rely on "Auto" mode to pick the most appropriate available model for any request should be extremely careful, as some users report it can switch to expensive models for extremely simple queries).
That's a big change from previous months, when GitHub Copilot subscribers were allocated a certain number of "requests" and "premium requests" based on their payment tier. GitHub said that the old system meant that "a quick chat question and a multi-hour autonomous coding session [could] cost the user the same amount," forcing Copilot itself to "absorb much of the escalating inference cost behind that usage." [...] Indeed, some Copilot users have been sharing estimates from GitHub's own tool showing that their previous monthly usage would rack up bills in the thousands of dollars under the new pricing plan. Under GitHub's new usage-based pricing system, paid Copilot subscriptions instead grant users a certain number of AI "credits" each month, with one credit corresponding to $0.01 of usage. Subscribers also get bonus credits depending on their subscription level: the $10/month Pro plan includes 1,500 credits ($15 worth); the $39 Pro+ plan includes 7,000 credits ($70 worth); and the $100/month Copilot Max plan includes 20,000 credits ($200 worth).
The precise number of Copilot credits used by a given prompt is determined by the number of input and output tokens used and the rates charged by the underlying large language model. That means pricing is highly dependent not just on the type of request but on the specific model that a user chooses. One million output tokens from OpenAI's GPT-5.4 nano would run just $1.25 on GitHub Copilot, but that same level of output would run $30 on the frontier GPT-5.5 model (Copilot users who rely on "Auto" mode to pick the most appropriate available model for any request should be extremely careful, as some users report it can switch to expensive models for extremely simple queries).
Should get really exciting. (Score:4, Interesting)
So long as they were losing money hand over fist the vendor does want to throw enough tokens at you to make you feel like you are having a good time; but as few as are required to do that since they lose money on every one. If they were breaking even or turning a profit the incentive would be to sneak as much spend and upsell in as possible; and it's well known that the verbosity/cost of LLM chatter is hard to predict; harder if there are multiple models and other complications being switched around in the background.
What sort of exciting little tricks will we see from vendors who actually make more if you use more?
Re: (Score:2)
new material? the reason ai porn is so prevalent is because it is pretty much the same for 100's of years if not longer.
Re: (Score:2)
So, this is in the process of being regulated as gambling?
Re: (Score:2)
True cost of AI LLMs (Score:3)
Hopefully AI thing will start to cool a bit now that the free ride is over and end users start to bear the full cost of these systems. Up until now the big AI companies have been giving it away at a token cost to keep the hype going.
Hopefully seeing the true costs will push demand more towards more efficient, smaller, specialized local models for many applications.
Re: (Score:3)
"Cool a bit"? If the general truth about the subsidizing of prices gets out, we are looking at a big bubble burst at least as bad as the dot-com poppage.
Investor funds and market-share-fights have kept AI prices low or free, but of course that can't last forever. I suspect one prominent but stressed AI company will spill the beans about fake pricing ("we all do it!"), putting pressure on the rest to prove that claim is false, which they'll fail, spooking investors, ending the run, and triggering a recession
Turn off automated PR reviews (Score:2)
I'm disabling all automatic PR reviews for sure. I was going to do this anyway, but it was a setting the org controlled so I didn't have a choice. Now I think we have an argument for ending Copilot's review feedback. Because while the models are fine at reviewing code, mostly, the Copilot tool itself is garbage.
Post Customer Acquisition (Score:3)
Now who is surprised that prices go up after the customer acquisition phase?
Memory and compute are expensive. Data centers are aren't as easy to build as once thought. [finance-commerce.com]. Power is expensive. People fights over land and water rights.
And if the rest of the pack like Claude [theregister.com] moving to usage based limits. Then might as well make the change, Microsoft isn't a startup which is burning investor cash. They are definitely going to make a profit. AI ain't cheap and they already have the customers.
Re: (Score:3)
Now who is surprised that prices go up after the customer acquisition phase?
Like all drug dealers, the first hit is always free.
They should have waited until more jobs were eliminated and companies were more dependent on them.
I switched to Cursor AI after Copilot rolled out the 300 request cap. It is vastly superior overall, and whatever type of account my company provides, I've yet to hit a usage limit.
Re: (Score:2)
So long and thanks for all the circular funding (Score:3)
Let's recap. First, VC and mega-tech corporate coffers were the sources of unlimited funding. Compute was bought and essentially offered free or nearly free to even the most voracious users to juice usage stats and hopefully disrupt that too-long-entrenched habit people have of, let me check my notes, "using their brain to think." And to do 30% of the job of a person at 5% of the cost, sure, why not! Four to six months ago, those funds started drying up.
Fortunately, there was circular AI "investments" and "deals" between Nvidia, OpenAI, Microsoft, etc, which of course wasn't funding at all, just moving money between multinationals/hyperscalers to double- or triple-count the value of the money as it changed hands and maintain the air of inevitability by sheer scale of AI investment. Sure, it wasn't real value, but it was good enough to (maybe, if SpaceX, OpenAI and Anthropic move quickly) give runway for a few IPOs, and heck, when we're only burning through a few billion dollars a month, what's the harm in placing a bet to own the all future labor value?
But now? It looks the CEOs, boards, and shareholders remember they like their present-day selves more than their future-selves (future me? That guy is basically a stranger!), who are really the beneficiaries of these costly bets. So, with no other funding sources, AI companies have no choice but to charge users for compute directly. So now AI can do 30% of the job of a person at 10,000% of the cost.
I think we all see where this is going.
Re: (Score:2)
that too-long-entrenched habit people have of, let me check my notes, "using their brain to think."
[citation needed]
Alternatives (Score:3)
If AI services are becoming too expensive in the current environment, we can look to nature for help. There is a an abundant species of large mammals in the ape family that can be trained to do this kind of work as well.
Re: (Score:3)
Reminds me of The Feeling of Power [wikipedia.org] by Isaac Asimov.
Re: (Score:2)
At this point a man called Gardrilla Manceframe rediscovered and patented a device he had seen in a history book called a staircase. It has been calculated that his most recent tax bill paid for the social security of five thousand redundant Sirius Cybernetics Workers, the hospitalisation of a hundred Sirius Cybernetics Executives, and the psychiatric treatment of over seventeen-and-a-half-thousand neurotic lifts
When confronted with the Sirius Cybernetics elevators ("Happy Vertical People Transporters").
Re: (Score:1)
But they smell like stale Cheese-Whiz
Re: (Score:2)
Dune's Mentats.
Re: (Score:2)
Even at the more accurate prices it's still cheaper than paying a six figure salary. So while we will probably see a lot less AI doing peoples' hobby projects, or even open source projects, companies are not going to reduce their internal use of it.
the point is (Score:1)
It costs far far more than $200 a month to use AI in the way people are doing. The people referred to in the article (I'm assuming you did not read it, in the fine tradition of Slashdot) are at something like $13k of tokens used per month. And since the AI is only an adjunct to a human brain, even if it doubled the throughput of an engineer it would not be worth it, and of course it does not double productivity.
AI is absolutely amazing for some things, even finding bugs ("the app crashed when I did X, I thi
Re: (Score:3)
Well, that's a shock. Not. (Score:4, Informative)
It's the classic enshittification cycle, also beloved by drug-dealers.
1. Get users hooked on free or cheap service.
2. Once they're hooked, jack up the price.
Re: (Score:2)
In the AI coding space there are many alternatives. And, as reported, this is a huge change- from $10/month to $10 a day. They're not boiling the frog - creeping prices up slowly so no one quite notices.. Watch for the unsubscribes - mine included...
Someone has to pay for it (Score:3)
Comments on here repeatedly say people should only pay for what they use, whether this be water or electricity. Using AI is no different. The more you use the more you pay.
Not sure why this is controversial or surprising.
How much per token? (Score:2)
The problem with all these pricing models is that the concept of a "token" and what it consists of is nebulous at best. I know what they mean by it, but there's not hard calculation to determine how many "tokens" a specific prompt or response will consume. You can guess, but only so accurately. That concept needs to be better defined before anyone can estimate their ongoing costs.