Transcript
00:00:00So that certainly was quick. Last week I was talking in an episode about the end of
00:00:07cheap tokens where I broke down the economics of how these AI companies make or lose money
00:00:15and the role of training versus inference and why they're incentivized to move away from these
00:00:21subsidized subscriptions now that we have these agentic long running tasks that consume a lot of
00:00:27tokens. And today I woke up or actually I got it yesterday evening, I got an email by GitHub
00:00:34that they are announcing that all GitHub copilot plans so individual and business and enterprise
00:00:43will transition to usage based billing on June 1st. And I find their system really weird,
00:00:52but I'll get back to that. But it is no surprise. Of course, as mentioned, I broke it down in that
00:00:57separate episode. And yeah, for these companies, these subscription offerings they have are just
00:01:04not viable for these agentic tasks. And I'll also get back to why GitHub moved first here and why I
00:01:13think that, for example, OpenAI and also probably Anthropic will wait a bit longer until they move
00:01:21away from the subscription offerings. At some point for the agentic coding, I'm pretty sure they will
00:01:27all move to usage based pricing, though. But let's stick to GitHub. Now, what changes here with GitHub
00:01:34copilot is that, as I mentioned in this blog post, as I mentioned in the last episode, we now have
00:01:42these long running, these multi hour autonomous coding sessions that burn a lot of tokens. And
00:01:49well, that's just not viable anymore for them. GitHub has absorbed much of the escalating
00:01:54inference cost behind that usage. But the current premium request model is no longer
00:01:59sustainable for them. That's the reason why they're moving away from the subscription offering. And
00:02:05what's changing is that now in the future, you can still have a subscription. And if you have one,
00:02:11it will not be canceled for you. But instead of getting some guaranteed amount of usage out of it,
00:02:19you'll get AI credits. And what you'll get is, for now, if you have a business plan,
00:02:30and also actually if you have an individual plan, you'll get $10 of credits for your $10 per month
00:02:38plan and $39 for the $39 per month plan. For businesses, it's the same but there during a
00:02:46transition period, which goes to the end of August, you will actually get $30 and $70 in monthly
00:02:54credits clearly with the goal of ensuring that less business customers are canceling their
00:03:00subscriptions. But of course, once that transition period here is over, as a business customer,
00:03:06you will also just get a one to one match, so to say. So you spend $19 per month for your
00:03:12subscription, you get $19 in monthly AI credits and therefore in the end, unless I'm getting
00:03:19something totally wrong here, this is really just a prepayment. So I don't see why you would choose
00:03:28that over just paying for the actual use instead of having a subscription because of course with
00:03:35this model the disadvantage is that if you for some reason don't use up your usage, you burn some
00:03:42money but you don't really get an advantage out of having a subscription unless the models you can use
00:03:49through the subscription are priced cheaper than if you're directly using them through an API
00:03:55key of course. Now we can take a look at their pricing page here and there we see the prices we
00:04:05pay for these different models and in the end that is what we'll be paying and what will be deducted
00:04:13from our AI credits once we are in that usage based pricing plan. Here we see that one AI credit is
00:04:20one US dollar cent and you get 1900 credits for the business plan, that's the $19 per month plan
00:04:31and 3900 for the more expensive $39 plan so that in the end comes out to these amounts I mentioned
00:04:38here. And then we have this pricing table and from everything I see here from everything I could find
00:04:44here that is exactly well that there is no advantage here there is no discount here if you have a
00:04:55subscription so if we take a look at the price here for example for GPD 5.5 $5 per million input tokens
00:05:02and $30 per million output tokens well that is exactly the same what we're paying in the end
00:05:09if we're using the model directly through the open AI API so if we would bring our own API key in the
00:05:16end right so that's what I mean I don't see where exactly we would have an advantage here from using
00:05:24that subscription at least when it comes to these open AI models but it's the same of course for
00:05:30the Claude models if I go to the Claude API docs I can see that for Opus 4.7 for input tokens I am
00:05:39paying five dollars per million and for output I'm at 25 and again if I go back to the GitHub co-pilot
00:05:45pricing page here that is five dollar input 25 per output so again I don't see which advantage
00:05:55this subscription has I can instead just go directly for usage based pricing by bringing my own API key
00:06:03for example and then I have more flexibility and if I'm using less tokens I'm at least not charged
00:06:12these 19 or 39 dollars per month but maybe I'm also getting something totally wrong here and for
00:06:18businesses of course there may be different incentives here to use such subscriptions over
00:06:25usage based pricing because of course one advantage you could see here is that if you go for these
00:06:31subscriptions by default if you exceed or if you reach the limit of your subscription so if you use
00:06:39up all the credits you have in that subscription you can't continue using GitHub co-pilot and of
00:06:45course from a company perspective that may be what you want so that you give your employees these
00:06:51subscriptions here and if they're out of credits they're out of credits and they have to go back
00:06:56to writing code like a caveman by hand so I guess that could be one advantage here versus if you
00:07:02bring your own API key there is unlimited use which of course is not entirely true you can limit usage
00:07:08per API key too in the dashboards in the settings of these API providers of these AI providers but
00:07:16yeah that is the offering we have here a bit weird but as I mentioned not surprising at all because
00:07:23it's all not sustainable at least for now for GitHub co-pilot so for Microsoft in the end and
00:07:30I don't think that OpenAI will change its subscription pricing or how it works in the very
00:07:39near future at least they will certainly change it at some point and so will Enthropic and I mentioned
00:07:44that in the other episode but not right now I think because OpenAI just missed their revenue targets
00:07:52and that's an article from today too and that is a statement by their CFO and this is important
00:07:59in that statement the CFO said that she's worried the company might not be able to pay for future
00:08:06computing contracts if revenue doesn't grow fast enough now why does this mean that they will keep
00:08:14on subsidizing subscriptions well they definitely still wanna gain market share obviously at some
00:08:21point they will need to cut their losses but of course for one if they can achieve that with their
00:08:27subscriptions they're at least break even or not losing too much money they can grow their market
00:08:36share through those subscription offerings and then of course once they do switch to a usage-based
00:08:41pricing they can probably turn a profit there at least on the individual level per user and that
00:08:49of course could then fund future investments the problem OpenAI of course has is that everybody
00:08:55knows chatgpt everybody's using chatgpt but 90 ish percent of all the users are not paying for it
00:09:04they're using the free version and of course well no surprise that isn't making any money for OpenAI
00:09:12and then of course you have all these subscribers that are using their subscription for codecs so
00:09:17for agentic coding and those subscribers are not super profitable that is the situation in which
00:09:23OpenAI finds itself but still as mentioned they need to grow their market share so that eventually
00:09:29they can make changes to the pricing get rid of subscriptions or make them more expensive or
00:09:34whatever limit the usage so that they can turn a profit per user and then fund future investments and
00:09:40therefore i think that right now where they are still desperate to grow their revenue for now it's
00:09:47the revenue focus because that equals market share in the end so now where they're desperate to grow
00:09:53their visibility their market share the amount of paying users i don't think they will make any
00:09:59changes to their codecs subscription plan offerings anytime soon or for now at least and for the same
00:10:05reason i think anthropic may not be making drastic changes i do very well see that anthropic may
00:10:13remove the more expensive models opus models from the base subscription plans but will keep it
00:10:19in the more expensive plans because the way the limits in those cloth subscription plans work
00:10:27they're already pretty restrictive not as bad in my experience as some people make it seem on x but
00:10:34definitely more restrictive than the open ai usage limits in the codec subscriptions so anthropic may
00:10:41not be losing as much money on these subscriptions and of course i'm pretty sure they don't want to
00:10:47give up market share and give that to open ai of course they're fierce competitors the advantage
00:10:54and i already mentioned that in the other episode too of course is that anthropic is very well
00:10:58positioned in the enterprise market so yeah they can at some point and probably before open ai does
00:11:06it switch to a more aggressive pricing model potentially to usage based pricing model just
00:11:12like github copilot did here so that they can also turn some profit make some profit out of these
00:11:18users and then fund future investments now for github and therefore microsoft in the end things
00:11:24are a bit different in my opinion because of course whilst they certainly still also want to gain
00:11:31market share it's probably not the number one priority for them because microsoft is already
00:11:37shoving their copilot offerings everybody's well is sharing it with everybody and is pushing it into
00:11:45all companies and into all their products and of course they are for naturally because so many
00:11:51companies are using microsoft products and are also using vs code or just visual studio so naturally
00:11:59github copilot will be used by a decent amount of companies if they want or if their users their
00:12:06employees want it or not so market share is probably not their number one priority and of
00:12:12course therefore they are definitely not interested in burning money long term they especially don't
00:12:19need to just gain market share to eventually become profitable because they have no training to be
00:12:26funded right they just need a profitable user base and as mentioned market share kind of comes on its
00:12:34own from a microsoft perspective because copilot is everywhere and they have that they have these
00:12:40deep roots in all these companies so that's why i think they are now more interested in cutting
00:12:45their losses and already starting to turn a profit per user or at least not lose money i think that is
00:12:52their goal here and therefore yeah github copilot was the first one to move to usage based pricing
00:13:00definitely won't be the last one but i think for the reasons mentioned codex and or open ai and
00:13:07frobic will probably not move to usage based pricing just now and frobic maybe
00:13:13but open ai i really don't see that happening in the very near future