Tokens are getting more expensive
ethanding.substack.com
ethanding.substack.com
> consumers hate metered billing. they'd rather overpay for unlimited than get surprised by a bill.
Yes and no.
Take Amazon. You think your costs are known and WHAMMO surprise bill. Why do you get a surprise bill? Because you cannot say 'Turn shit off at X money per month'. Can't do it. Not an option.
All of these 'Surprise Net 30' offerings are the same. You think you're getting a stable price until GOTAHCA.
Now, metered billing can actually be good, when the user knows exactly where they stand on the metering AND can set maximums so their budget doesn't go over.
Taken realistically, as an AI company, you provide a 'used tokens/total tokens' bar graph, tokens per response, and estimated amount of responses before exceeding.
Again, don't surprise the user. But that's an anathema to companies who want to hide tokens to dollars, the same way gambling companies obfuscate 'corporate bux' to USD.
It’s nearly impossible to tell what the hell is going where and we are mostly surviving on enterprise discounts from negotiations.
The worst thing is they worked out you can blend costs in using AWS marketplace without having to raise due diligence on a new vendor or PO. So up it goes even more.
Not my department or funeral fortunately. Our AWS account is about $15 a month.
The way to deal with this is with an org-level Service Control Policy that enforces the tagging standards.
A resource doesn't have the right tags associated with it? It can't be created.
https://docs.aws.amazon.com/organizations/latest/userguide/o...
Not a bug, a feature.
I am not saying this is desirable, but it is necessary IFF you chose to use these services. They are complex by design, and intended primarily for large scale users who do have the expertise to handle the complexity.
Pricing schemes like these just make them move back to virtual machines with "unlimited" shared cpu usage and setting up services (db,...) manually.
You could also have potential customers who would be interested in your solution, but don't want it hosted by an American company. Spinning up a few Hetzner VMs is easy. Finding European alternatives to all the different "serverless" services Amazon offers is hard.
Not happened yet. The nearest I have come to it was a requirement that certain medical information stays in the UK, and that is satisfied by using AWS (or other American suppliers) as long as its hosted in the UK.
Most small business I have dealt with use AWS do just need a VPS. If they are willing to move to a scary unknown supplier I suggest (unknown to them, very often one that would be well known to people on HN) then I suggest AWS Lightsail which is pretty much a normal VPS with VPS pricing - it significantly cheaper than an instance plus storage, just from buying them bundled (which, to be fair to Amazon, is common practice).
My own stuff goes on VPSs.
Except it is still Amazon and subject to the same weird billing practices. I once terminated a Lightsail instance and they kept charging me, claiming that I didn't terminate the static IP address associated with it. The IP address itself cost the same as the instance + IP address did.
Now, that would make sense in "real" AWS, but you'd expect it to be more straightforward with a simplified service like Lightsail.
At that point wouldn't it simply be cheaper to do VMs?
I think a lot of people are missing a key part of the wording of my comment, that capitalised for emphasis "IFF" (which means "if and only if").
I am absolutely certain a lot of people would save money using VMs - or at scale bare metal.
IMO a lot of people are using AWS because it is a "safe" choice management buy into that is not expensive in context (its not a big proportion of costs).
The point where you get sticker shock from AWS is often significantly lower than the point where you have enough money to hire in either of those roles. AWS is obviously the infrastructure of choice if you plan to scale. The problem is that scaling on expertise isn’t instant and that’s where you’re more likely to make a careless mistake and deploy something relatively costly.
This:
> The point where you get sticker shock from AWS is often significantly lower than the point where you have enough money to hire in either of those roles
makes me doubt this:
> AWS is obviously the infrastructure of choice if you plan to scale.
If you can afford the large fixed cost of vertical integration, it's always cheaper to do things yourself, so the sweet spot for using providers like AWS is scaling down, not up. A managed DB lets you hire a fraction of a sysadmin or devops person from AWS.
The moment you end up paying enough to AWS to hire a sysadmin, you basically are getting an antagonistic sysadmin from AWS, whose primary goal is to make as much money off you as possible. The incentives are not aligned.
What a baffling comment. Is it normal to even consider hiring someone to figure out how you are being billed by a service? You started with one problem and now you have at least two? And what kind of perverse incentive are you creating? Don't you think your "finops" person has a vested interest in preserving their job by ensuring billing complexity will always be there?
Is it, though? At best someone wearing that hat will explain the bill you're getting. What value do you get from that?
To cut costs, either you microoptimize things, of you redesign systems to shed expenses. The former gets you nothing, the latter is not something a "finops" (whatever that is supposed to mean) brings to the table.
I did say it applies IFF and only IFF you choose to use these services, and if you have chosen to use these services you have presumably decided they are good value for money. If not, why are they using AWS.
Of course the complexity and extra cost of managing the billing is something that someone who has chosen to use AWS has already factored in, right?
The alternative is to not use AWS.
If and only if and only if and only if? :)
(also, while on the topic, I think a simple "if" covers it here, since the relationship is not bidirectional)
Absolutely. This was common for complicated services like telecom/long distance even in the pre-cloud days. Big companies would have a staff or hire a service to review telecom bills and make sure they weren’t overpaying.
If I charged compute based on the number of micro-ops executed then that would be a clear definition, but the actual cost would not be something you could predict, as it would depend on what architecture of CPU you ended up running it on.
AWS is even more complicated and variable than that as for cloud storage you have to deal with not only the costs of the different storage classes, but also early deletion fees, access charges, etc. Combined it makes it impractical to work out how much deleting a file from cloud storage will save (or cost). Sure you could probably calculate it if you knew the entire billing history of the file and the bucket it is in, but do you really want to do that every time you delete a file?
While I don't know enough to say if this is intentional, as it could result from simply blindly optimizing for profit, this sort of pricing model is anti-capitalistic as it prevents consumers from truly making informed decisions. We see the same thing is the US healthcare system, where no one can actually tell you how much an operation will cost ahead of time. That creates a very inefficient (but very profitable) market.
I made it clear that you ask the user to choose between 'accept risk of overrun and keep running stuff', 'shut down all stuff on exceeding $ number', or even a 'shut down these services on exceeding number', or other possible ways to limit and control costs.
The cloud companies do not want to permit this because they would lose money over surprise billing.
Up until recently, you could hit somebody else's S3 endpoint, no auth, and get 403's that would charge them 10s of thousands of dollars. Coudnt even firewall it. And no way to see, or anything. Number go up every 15-30 minutes in cost dashboard.
Real responsibility is 'I have 100$ a month for cloud compute'. Give me a easy way to view it, and shut down if I exceed that. That's real responsibility, that Scamazon, Azure, Google - none of them 'permit'.
They (and well, you) instead say "you can build some shitty clone of the functionality we should have provided, but we would make less money".
Oh, and your lambda job? That too costs money. It should not cost more money to detect and stop stuff on 'too much cost' report.
This should be a default feature of cloud: uncapped costs, or stop services
Perhaps requiring support for bill capping is the right way to go, but honestly I don’t see why providers don’t compete at all here. Customers would flock to any platform with something like “You set a budget and uptime requirements, we’ll figure out what needs to be done”, with some sort of managed auto-adjustment and a guarantee of no overage charges.
Ah well, one can only dream.
Because the types of customers that make them the most money don't care about any of this stuff. They'll happily pay whatever AWS (or other cloud provider) charges them, either because "scale" or because the decision makers don't realize there are better options for them. (And depending on the use case, sometimes there aren't.)
Then the goal would be to set the resource limits to something you are happy with.
Yes, this is a pain in the ass to set up and AWS will probably never implement this, but it is the correct solution.
This is the exact same thing that frustrates me with GitHub's AI rollout. Been trialing the new Copilot agent, and it's cost is fully opaque. Multiple references to "premium requests" that don't show up real-time in my dashboard, not clear how many I have in total/left, and when these premium requests are referenced in the UI they link to the documentation that also doesn't talk about limits (instead of linking to the associated billing dashboard).
* One chat message -> one premium credit (most at 1 credit but some are less and some, like opus, are 10x)
* Edit mode is the same as Ask/chat
* One agent session (meaning you start a new agent chat) is one "request" so you can have multiple messages and they cost the credit cost of one chat message.
300 or 1500 per month depending on plan. $0.04 per premium request i believe.
This is coding agent, the asynchronous copilot, not the agent chatmode in copilot plugins for vscode etc
But for users, that fine grained cost is not good, because you’re forcing a user to be accountable with metrics that aren’t tied to their productivity. When I was an intern in the 90s, I was at a company that required approval to make long distance phone calls. Some bureaucrat would assess whether my 20 minute phone call was justified and could charge me if my monthly expense was over some limit. Not fun.
Flat rate is the way to go for user ai, until you understand the value in the business and the providers start looking for margin. If I make a $40/hr analyst 20% more productive, that’s worth $16k of value - the $200/mo ChatGPT Pro is a steal.
But that means that if you were conned into using infrastructure that actually costs more than the alternative, making your cost structure worse, you're still going to eat the loss because it's not worth taking your devs time to switch back.
But tokens don't quite have this problem -yet. Most of us can still do development the old way, and it's not a project to turn it off. Expect this to change though.
But for AI in the context of point-solutions and on-the-job use cases, metered billing is a death blow.
In this context, metered is a massive incentive to not use the product and requires the huge friction of having to do a cost/benefit analysis before every task. And if you're using it at work you may even need management sign-off before you can use it again.
For a tool that's intended to amplify productivity, very few humans want to make a cost/benefit analysis 250 times a day whether it's worth $3 to code up a boilerplate or not. On metered billing, they just wont use it.
That's what we're supposed to do, right?
So let's see if we can spend a few tokens to ask the LLM for a cost/benefit analysis of using an LLM to solve the problem. I'd bet we can trust the result...
Adam Smith of all people writes of this:
By necessaries I understand not only the commodities which are indispensably necessary for the support of life, but whatever the custom of the country renders it indecent for creditable people, even of the lowest order, to be without. A linen shirt, for example, is, strictly speaking, not a necessary of life. The Greeks and Romans lived, I suppose, very comfortably though they had no linen. But in the present times, through the greater part of Europe, a creditable day-labourer would be ashamed to appear in public without a linen shirt, the want of which would be supposed to denote that disgraceful degree of poverty which, it is presumed, nobody can well fall into without extreme bad conduct. Custom, in the same manner, has rendered leather shoes a necessary of life in England. The poorest creditable person of either sex would be ashamed to appear in public without them.
<https://en.wikisource.org/wiki/The_Wealth_of_Nations/Book_V/...>
Running water, sewerage, gas service, home postal service, electricity, automobiles, telephone service, Internet service, mobile phones, universal healthcare, and many other utilities were once considered luxuries but came to be recognised as essentials.
Jokes aside, I already don't live without my LLM. Rely on it too much. Though more for comfort than survival.
Fair points, I'll add that I'm not sold that LLMs will become indispensable. There's plenty of flashy tech which hasn't.
But of the past 20--30 years' or so offerings, it seems to have a reasonably high chance.
Asking “what day is today” vs “create this api endpoint to adjust the inventory” will cost vastly different. And honestly I have no clue where to start to even estimate the cost unless I run the query.
Budget Billing is a free program that helps you easily manage your monthly energy costs . We calculate your monthly payment amount based on your average energy costs over the last 12 months and adjust your payment amount each month, so that you don't have big spikes on your bill.
While it's not a savings program, Budget Billing helps you stay in control of your bill by avoiding seasonal bill spikes.
<https://www.pge.com/en/account/billing-and-assistance/financ...>
It's not at all flat rate, though, as you'll settle the balance after the year like a tax filing (refund or back pay).
I assume the utility support would explain this way of calculating/predicting/tracking if one asks nicely.
The reasons for not smoothing out further than monthly payment intervals in a yearly measurement interval is AFAIK due to regulatory restrictions on credit/debt, and because this is more due to a metering process efficciency/overhead limitation than because they inherently want to offer these fixed/steady payments: you don't want to send out a meter reader more than about once a year, and doing it that often also conveniently spreads out seasonal variations letting the bill predictions(/estimates used for calculating the monthly payments that are due) work much nicer and easier to understand than some summer/winter split tracking.
Note this information is all about the default/baseline tariffs, which the local electrical energy seller (which is a distinct entity from the local electrical power/grid provider, who is the one that has the natural monopoly) can't refuse due to credit score or the like and which can't have any remotely long term/notice periods. Many tariffs are quite similar in style, though, due to the inherent limitations of yearly-read non-smart energy consumption counters.
Around now seems to be where finally you as a customer can demand upgrade to a smart meter which notably (from my understanding) gains the ability to live-report energy flux for 15 minute quantization/intervals, matching the European super grid's electricity market. A nice side-effect closely related to that market quantization is that these meters allow selling PV production to other units/properties nearby (at least where the connection/current path doesn't actually go through grid operator owned wiring, e.g. units in an apartment building) without having to pay grid fees on that energy sale despite the energy going through the meters instead of being sold through a behind-the-meter bypass feed/line. That "virtual" metering of the hyper-local PV energy sale makes it just much easier in terms of infrastructure, compared to some actual real-time load-balancing setup that somehow spreads an apartment building's monolithic PV array's power across multiple units's behind-the-meter breakerboxes taking care to not wastefully oversupply any one as backfeed would at best be compensated at about a third of the all-fees-included retail rate (it's about the same as the grid fee part that's payable by non-local buyers of electricity, both around 10ct/kWh) instead of 1:1 compensating the approximately 30ct/kWh buy-rate of some local apartment's meter. Also, less active infrastructure at a scale that doesn't afford on-site 24/7 staffing for monitoring and customer-side-uptime-enhancing redundancy features is always good. I'm counting a mere meter that only meters and doesn't steer/control electricity as passive infrastructure from the power flow's POV. Especially if it's built to fail in a way where power still flows but the metering/counting is broken, at least where there's more than just yearly reporting from the meter (and thus non-metering would be caught quickly).
Electricity uptime of this grade is something with perhaps underappreciated benefits: pretty much all but critical systems and those that run non-crash-safe software will not require a UPS, as power outages at least in urban areas (except those where local distribution systems drown in climate change flash flooding) happen once every few years, either due to larger-scale blackouts (entire city/metro/state) or from non-redundant local infrastructure dying of old age.
So no in fact you do get X tokens for 20 dollars and off your usage is too high they would absolutely put a stop to that. It wouldn’t even surprise me if they have per customer rate limits.
Where is the distinction here, it isn’t an unlimited plan, the limit just isn’t communicated.
"Texas freeze raises concerns about ‘ridiculous’ variable rate bills" (2021)
HOUSTON, Feb 23 (Reuters) - In Spring, Texas, about 20 miles (32 km) north of Houston, Akilah Scott-Amos is staring down a more than $11,000 electric bill for this month, a far cry from her $34 bill at this time last year.
"What am I going to do?" Scott-Amos, 43, said. She was among the millions of Texas residents who lost power during several days of bitter cold that caused the state's electrical grid, operated by the Electric Reliability Council of Texas, to break down. "I guess the option is, what, I'll pay it? I just don't feel like we should have to."
<https://www.reuters.com/business/energy/texas-freeze-raises-...>
This instance involves variable rate billing systems, where the cost per unit use can vary tremendously. There are also stories you'll find where equipment issues (broken water or gas mains, malfunctioning electrical equipment, failure to terminate per-minute-billed telephone calls, etc.). I can remember the latter featuring in some compendium of records (possibly Guiness) for a family who hadn't realised that they needed to hang up the phone between calls and got hit with a staggering (at least for the time) bill. More recently, data and roaming charges tend to be culprits:
<https://worthly.com/most-expensive/expensive-phone-bills-tim...>
In all of these cases, a significant problem is that usage is all but entirely divorced from metering, and people have no idea of what their actual usage patterns are until "invoice therapy" arrives.
(A term I'm borrowing from a friend, though I'll note that monthly feedback cycles on what are typically hourly/daily decisionmaking patterns proves a poor systems control theory match.)
Nor do they have stable don't-give-me-any-better "answer quality" expectations from models. At least not yet.
AI model services don't have any version of the typical utility's geographical market protection or monopoly.
Worse, they have all been funded for massive loss-leader hyper growth, so to survive in the short term, they have to compete on who can raise and lose more.
Then looking into the long term, unlike the Uber's of the world, the point where a market will reach saturation, survivors finally "win", and margins can transition from negative to positive, keeps moving away.
I don't see any way in which utility bills are economically comparable to this situation.
The only "hope" the general AI service business has, is that at some point one of them pulls far enough ahead of the others that everyone else's economics collapse first.
Like a very high speed version of the centralization in advanced chip fabs. But with no profitability until the centralization happens. Brutal.
Which is what TFA describes, and is why there is no unlimited flat-rate deal for utilities. But that's a mature market that is not relying on growth for valuations, and isn't appealing to VCs trained on net/mobile/crypto bubbles.
But those contracts are so much more expensive that virtually no one gets them.
I can't see any situation where true unlimited usage for a flat rate, even a wildly expensive and uneconomical one, would make any sense because as often happens you'd basically be incentivizing people to waste as much gas, electricity and water as they possibly can to "get their money's worth" or whatever, and we should be encouraging the exact opposite for those 3 things.
Power in Greece works the same way, except you (are supposed to) get the guess one month and the count the next (and pay the difference). In practice, they count less often than that.
We own an old 170 sq.m. semi-detached house, and the electricity bill for June was 20 euros. That's with all heating & hot water coming from electricity (heat pumps) and owning an EV that we charge at home.
That's not the motive where I live. There's basically 2 reasons: (1) customers (esp. those, often less well off, that resort to “money earned, money spent”) like the predictability of it; and (2) depending on the season spending changes significantly (heating, cooling), whereas salaries don't.
Usually the corrective amount is also paid out over a period of months if it exceeds X% of one month's spend.
Although here, if people get into debt they can be forced onto pre-payment meters where that's not an option, and the unit price is also higher. There's a lot of controversy about this.
In manufacturing, everything is about consistency.
On the web, it’s not clear that one website is being massively more wasteful than another. From your perspective you’re engaging in the same behaviour (opening a website) but for reasons outside your control you may be spending more than intended.
I remember not having unlimited internet traffic at home and always having to worry. These days, I pay a flat monthly fee for internet access.
The point is that infrastructure is non-optional, you're not going to decide not to use it, and you rarely switch (in the case of utilities you often CAN'T switch). You set it and forget it.
AI in the context of an end-user product or tool is not that (unless you're running the AI product company, where it becomes infra for you).
It's a blender or a kitchenaid mixer. Let's say companies gave those away for free but only charged metered fees to use them. If you had to pay $4 to blend something or run your mixer (and weren't sure if the tool would get it right the first time or you'd have to re-do it 5 times) you'd use them much less. They are optional.
You'd treat the decision to bake like a financial one. Nobody asks their spouse if it's okay to take a shower.
It's great for free experimentation when coding apps that use API inference, it's actually motivating to build uses for it because it feels akin to a benefit that otherwise goes unused. Of course there is some price risk in holding their crypto token. After a large spike and drop in the first couple days, it's held steady for 6 months with a $2-$3 floor, possibly due to it having that baseline utility. Mentally, it's far more comfortable to me to stake a principal sum that I can wholly unstake in the future, than to spend on per-request fees.
Like, lets have a real talk, shall we? Lets just assume that the topic we are discussing on that I am right and you are wrong, How can I even convince ya when you are showing so less of maturity...
And lets say that you are right and I am wrong, but the fact that you are being so bullish on the fact that you can't be wrong and bringing the "this is worse than reddit" and etc. can't make me take you serious and can't make me think your opinion is valid.
If you really want, I'd like to logically disect this stuff as adults using pure logic & not mere opinions.
Waiting for your response.
Standard packages are like insurance. Everyone pays more or less the same premium, but some claim more than others. On average people always overpay for insurance.
The upside is that it's a predictable cost for the users, and also means predictable cash flow for the provider.
You get a surprise bill because of surprise usage of services billed based on usage.
If you ask anyone how much water or electricity they use per month, the first thing they're going to do is look at last month's usage.
Estimating what you need ahead of time is a hard problem.
In fairness, AWS doesn't give you a lot of tools to help you measure and predict how many "units" you'll use outside of running your thing and measuring. On the other hand, running your thing and measuring is the defacto way to figure out how much of something you'll use.
Finally, there are AWS services like EC2 and RDS you can run at a fixed cost to help you stay within budget. Traffic/bandwidth is the only thing that comes to mind that you're pretty much required to use without a way to fix the cost (although you can get pretty close with bandwidth limits on EC2 interfaces)
> when a new model is released as the SOTA, 99% of the demand immediately shifts over to it
99% is in the wrong ballpark. Lots of users use Sonnet 4 over Opus 4, despite Opus being 'more' SOTA. Lots of users use 4o over o3 or Gemini over Claude. In fact it's never been a closer race on who is the 'best': https://openrouter.ai/rankings
>switch from opus ($75/m tokens) to sonnet ($15/m) when things get heavy. optimize with haiku for reading. like aws autoscaling, but for brains.
they almost certainly built this behavior directly into the model weights
???
Overall the article seems to argue that companies are running into issues with usage-based pricing due to consumers not accepting or being used to usage based pricing and it's difficult to be the first person to crack and switch to usage based.
I don't think it's as big of an issue as the author makes it out to be. We've seen this play out before in cloud hosting.
- Lots of consumers are OK with a flat fee per month and using an inferior model. 4o is objectively inferior to o3 but millions of people use it (or don't know any better). The free ChatGPT is even worse than 4o and the vast majority of chatgpt visitors use it!
- Heavy users or businesses consume via API and usage based pricing (see cloud). This is almost certainly profitable.
- Fundamentally most of these startups are B2B, not B2C
Thank you for pointing out that fact. Sometimes it's very hard to keep perspective.
Sometimes I use Mistral as my main LLM. I know it's not lauded as the top performing LLM but the truth of the matter is that it's results are just as useful as the best models that ChatGPT/Gemini/Claude outputs, and it is way faster.
There is indeed diminished returns on the current blend of commercial LLMs. Deep seek already proved that cost can be a major factor and quality can even improve. I think we're very close to see competition based on price, which might be the reason there is so much talk about mixture of experts approaches and how specialized models can drive down cost while improving targeted output.
It's great if you can leave it unattended, but personally, coding's an active thing for me, and watching it go is really frustrating.
How much of that is the naming?
Personally I just avoid OpenAIs models entirely because I have absolutely no way of telling how their products stack up against one another or which to use for what. In what world does o3 sort higher than 4o?
If I have to research your products by name to determine what to use for something that is already a commodity, you've already lost and are ruled out.
There's also 4o-mini and o4-mini...
The article repeats this throughout but isn't it a straight lie? The plan was named 20x because it's 20x usage limits, it always had enforced 5 hour session limits, it always had (unenforced? soft?) 50 session per month limits.
It was limited, but not enough and very very probably still isn't, judging by my own usage. So I don't think the argument would even suffer from telling the truth.
I can’t believe how many comments and articles I’ve read that assume it was unlimited.
It’s like it has been repeated so many times that it’s assumed to be true.
Not every problem needs a SOTA generalist model, and as we get systems/services that are more "bundles" of different models with specific purposes I think we will see better usage graphs.
But we're still in the hype phase, people will come to their senses once the large model performance starts to plateau
Like what? People always talk about how amazing it is that they can run models on their own devices, but rarely mention what they actually use them for. For most use cases, small local models will always perform significantly worse than even the most inexpensive cloud models like Gemini Flash.
This shouldn't be that expensive even for large prompts since input is cheaper due to parallel processing.
In the food industry is it more profitable to sell whole cakes or just the sweetener?
The article makes a great point about replit and legacy ERP systems. The generative in generative AI will not replace storage, storage is where the margins live.
Unless the C in CRUD can eventually replace the R and U, with the D a no-op.
I really don't understand where you are trying to get. But on that example, cakes have a higher profit margin, and sweeteners have larger scale.
AI companies advertise peak AI performance, users select AI tools on worst case AI fuckups: hence, only SOTA is ever in demand. TFA illustrates this well.
AI will be judged on it's worst performance, just like people are fired for their worst showing, not their best. No one cares about AI performance in ideal (read: carefully contrived) settings. We care how bad it fucks up when we take our eyes off it for 2 seconds.
It's the same as compute--you can skip testing and throw money at the problem but you're going to end up paying more.
We have some pretty basic guidelines at work and I think that's a decent starting point. They amount to a few example prompts/problem types and which OpenAI model to try using first for best bang for your buck.
I think some of it also comes down to scale. Buying a 5 pack of sledgehammers isn't a terrible value when everything comes in a "5 pack" and you only need <= 5 tools total. Or more practically, on the small end it's more economical to run general purpose models than tailor more specific models. Once you start invoking them enough, there's a break even and flip point where spending more time on the tailored or custom model is cheaper.
This has been working great for the occasional use, I'd probably top up my account by $10 every few months. I figured the amount of tokens I use is vastly smaller than the packaged plans so it made sense to go with the cheaper, pay-as-you-go approach.
But since I've started dabbling in tooling like Claude Code, hoo-boy those tokens burn _fast_, like really fast. Yesterday I somehow burned through $5 of tokens in the space of about 15 minutes. I mean, sure, the Code tool is vastly different to asking an LLM about a certain topic, but I wasn't expecting such a huge leap, a lot of the token usage is masked from you I guess wrapped up in the ever increasing context + back/forth tool orchestration, but still
Everyone complains about the prices of other models but there are much cheaper alternatives out there and DS is no slouch either.
This seems like such an obvious idea that I'm sure everyone is already working on it!
Aider has an option where you can combine different models, so one does the thinking and one implements the changes.
https://aider.chat/docs/leaderboards/
e.g. o3 (high) + 4.1
There used to be more such combinations in this list but it seems they were removed. A while ago I think the best thing on the list was such a hybrid.
The meaningful frontier isn't scalar on just the capability, it's on capability for a given cost. The highest capability models are not where 99% of the demand is on. Actually the opposite.
To get an idea of what point on the frontier people prefer, have a look at the OpenRouter statistics (https://openrouter.ai/rankings). Claude Opus 4 has about 1% of their total usage, not 99%. Claude Sonnet 4 is the single most popular model at about 18%. The runners up in volume are Gemini Flash 2.0 and 2.5, which are in turn significantly cheaper than Sonnet 4.
One of the graphs even lists a "Claude 3.5 Opus", which does not exist. After 3.5 Sonnet was released, 3 Opus largely fell into irrelevance until they decided to finally release another big, expensive model with Opus 4, which still isn't anywhere near as popular as Sonnet 4 with users who pay API prices.
Second, why are SV people obsessed with fake exponentials? It's very clear that AI progress has only been exponential in the sense that people are throwing a lot more resources at AI then they did a couple years ago.
Is it done like this just to show it wasn't written by a LLM?
Thou needst to live in the archaic.
I don't agree with the Cognition conclusion either. Enterprises are fighting super hard to not have a long term buying contract when they know SOTA (app or model) is different every 6 months. They are keeping their switching costs low and making sure they own the workflow, not the tool. This is even more prominent after Slack restricted API usage for enterprise customers.
Making money on the infra is possible, but that again misunderstands the pricing power of Anthropic. Lovable, Replit etc. work because of Claude. Openai had codex, google had jules, both aren't as good in terms of taste compared to Claude. It's not the cli form factor which people love, it's the outcome they like. When Anthropic sees the money being left on the table in infra play, they will offer the same (at presumably better rates given Amazon is an investor) and likely repeat this strategy. Abstraction is a good play, only if you abstract it to the maximum possible levels.
On top of this Gemini CLI still doesn’t support paying through the Google AI subscription. I assume it’s some sort of bureaucratic reason that’s preventing them from moving quickly.
I'd prefer it just specify a number of tokens rather than be variable on demand - I see that lets them be more generous during low periods but the opacity of it all sucks. I have 5-minute time-of-use pricing on my electricty and can look up the current rate on my phone in an insant - why not simply provide an API to look up the current "demand factor" for Claude (along with the rules for how the demand factor can change - min and max values for example) and let it be fully transparent?
On one of the systems I'm developing I'm using LLMs to compile user intents to a DSL, without every looking at the real data to be examined. There are ways; increased context length is bad for speed, cost and scalability.
Even before AI, it felt like the value of intelligence and knowledge had been dropping over time. This makes sense as the internet has democratized access to information and promoted intellectual self-improvement. The supply of intelligence increased dramatically but demand for it struggled to keep up (in spite of the tech boom). Now demand for intelligence has plateaued; This is one way to look at current tech layoffs.
It got to a point that intelligence is almost worthless now. 'Earning' money is mostly about social connections, not intelligence. So all these use cases which people are using AI for are pointless in terms of earning money in the current system. The current system rewards money-acquisition, it does not reward value-creation. You don't need intelligence to acquire money in this system; you need social connections. AI does not give you social connections; if anything, it takes away social connections. The people using AI to build themselves an amazing second internet will have nobody to share it with; no users, no investment.
The oversupply of intelligence means that it cannot find any serious avenues to earn a financial return, so instead it turns to political manipulations because system reform (or manipulation) is the shortest path to earning monetary returns... Though often this manipulation only further decouples value creation from money-acquisition.
Actually when doing my first attempt at vibe coding a few months ago, I found that Gemini Flash was fine for my tasks, and way faster than the heavier models. So I found the smaller model a vastly superior user experience.
The speed really adds up when you're using the autonomous coding agents, since they tend to require many LLM calls for a few simple changes.
I managed a deep learning team at Capital One and the lock-in thing is real. Replit is an interesting case study for me because after a one week free agent trial I signed up for a one year subscription, had fun the their agent LLM-based coding assistant for a few weeks, and almost never used their coding agent after that, but I still have fun with Replit as an easy way to spin up Nix based coding environments. Replit seems to offer something for everyone.
Does this mean that other languages might offer better information density per token? And does this mean that we could invent a language that’s more efficient for these purposes, and something humans (perhaps only those who want a job as a prompt engineer) could be taught?
Kevin speak good? https://youtu.be/_K-L9uhsBLM?si=t3zuEAmspuvmefwz
https://www.science.org/content/article/human-speech-may-hav...
Granted English is probably going to have better quality output based on training data size
English (And any of the dominant languages that you could use in it's place) work significantly better than other languages purely by having significantly larger bodies of work for the LLM to work from
Maybe even something anyone can read and maybe write… so… Kevin English.
Job applications will ask for how well one can read and write Kevin.
Regarding cost per token: is a token ideally a composable, atomic unit of information? Since English is often used as an encoding format, efficiency is limited by English's encoding capacity.
Could other languages offer higher information density per token? Could a more efficient language be invented for this purpose, one teachable to humans, especially aspiring prompt engineers?
67 tokens vs 106 for the original.
Many languages don't have articles, you could probably strip them from this and still understand what it's saying.
In practice, the problem is that any such constructed language wouldn't have a corpus large enough to train on.
It's really unfortunate that we ended up with English as the global lingua franca right at the time generative AI came about, because it is effectively cementing that dominance. Even Chinese models are trained mostly on English AFAIK.
And the interesting property of Lojban is that it has unambiguous grammar that can be syntax-checked by tools and enforced by schemas, and machine-translated back to English. I experimented with it a bit and found that large SOTA models can generate reasonably accurate translations if you give them tools like dictionary and parser and tell them to iterate until they get a syntactically valid translation that parses into what they meant to say. So perhaps there is a way to generate a large enough dataset to train a model on; I wish I had enough $$$ to try this on a lark.
Edit: It says on the Jetbrains website:
“The AI Assistant plugin is not bundled and is not enabled in IntelliJ IDEA by default. AI Assistant will not be active and will not have access to your code unless you install the plugin, acquire a JetBrains AI Service license and give your explicit consent to JetBrains AI Terms of Service and JetBrains AI Acceptable Use Policy while installing the plugin.”
They didn’t cancel my existing ‘AI Pro’ subscription though, and have just let it keep running with no refunds.
Thanks, Jetbrains. You get worse every day.
It is standard practice with some coding agents to have different models for different tasks, like building and planning.
Look at Uber. Lost money for over a decade buying market share with venture capital. Now post IPO they have settled in to a position in the minds of users that’s hard to shake even as cheaper competition arrives. They have a durable business and a steady (even if not amazing) stock price.
And everything, I mean everything after the title is only a downhill:
> saying "this car is so much cheaper now!" while pointing at a 1995 honda civic misses the point. sure, that specific car is cheaper. but the 2025 toyota camry MSRPs at $30K.
Cars got cheaper. The only reason you don't feel it is trade barrier that stops BYD from flooding your local dealers.
> charge 10x the price point > $200/month when cursor charges $20. start with more buffer before the bleeding begins.
What does this even mean? The cheapest Cursor plan is $20, just like Claude Code. And the most expensive Cursor plan is $200, just like Claude Code. So clearly they're at the exact same price point.
> switch from opus ($75/m tokens) to sonnet ($15/m) when things get heavy. optimize with haiku for reading. like aws autoscaling, but for brains.
> they almost certainly built this behavior directly into the model weights, which is a paradigm shift we’ll probably see a lot more of
"I don't know how Claude built their models and I have no insider knowledge, but I have very strong opinions."
> 3. offload processing to user machines
What?
> ten. billion. tokens. that's 12,500 copies of war and peace. in a month.
Unironically quoting data from viberank leaderboard, which is just user-submitted number...
> it's that there is no flat subscription price that works in this new world.
The author doesn't know what throttling is...?
I've stopped reading here. I should've just closed the tab when I saw the first letter in each sentence isn't capitalized. This is so far the most glaring signal of slop. More than the overuse of em-dash and lists.
> when I saw the first letter in each sentence isn't capitalized. This is so far the most glaring signal of slop.
How so? It's the exact opposite imho. Lowercase everything with a staccato writing style to differentiate from AI slop, because LLMs usually don't write lowercase.
This comes across as sloppily written, but not sloppily generated.
just fyi, it's a very common manner of writing for younger folk online. more so in informal contexts, but as with everything else, once it's widely adopted it starts to creep into the more formal communication. it's not about "slop", it's just a cultural convention.
i should also note that many languages that got their orthographies defined relatively recently (e.g. various native american languages) use all-lowercase as well, by design. so there's no inherent reason why english can't do that either.
Which then might lead to you using a lot more, because it offsets some other thing that costs even more still, like your time.
With the primary advancements over the past two years being Chain Of Thought which absolutely obliterates token counts in what world would the "per token" value of a model be going up...
We need another attention is all you need, or three, imo. I hope the gold rush isn't impacting research, but I bet it is.
Sure I do!
I will consistently pick the fastest and cheapest model that will do the job.
Sonnet > Opus when coding
Haiku > Sonnet when fusing kitchen recipes, or answering questions where search results deliver the bulk of the value, and the LLM part is really just for summarizing.
You definitely want that for some tasks, but for the majority of tasks there is a lot of space for cheap & cheerful (and non-thinking)
They can deliver pretty much whatever they feel like. Who can tell a trash token from an hallucination? And tracking token usage is a pita.
Sum it up and it translates to: sell whatever you feel like at whatever price you feel like.
Nice!
Scaling laws let you spend more transistors and watts on intelligence
Do you want more tokens or smarter tokens?
brute force
you shouldn't be pricing compute directly (by charging for tokens yourself)
This extension might make the internet more accessible for you!
If you don't care to trivially make your text readable, then we for sure don't care to spend time to struggle through your text to see if there is any useful substance there.
And, I know this seems dramatic, but besides being cognitively distracting, it also makes me feel sad. Chatroom formatting in published writings is clearly a developing trend at this point, and I love my language so much. Not in a linguistic capacity - I'm not an English expert or anything, nor do I follow every rule - I mean in an emotional capacity.
I'm not trying to be condescending. This is a style choice, not "bad writing" in the typical sense. I realize there is often a lot of low-quality bitterness on both sides about this kind of thing.
Edit:
I also fear that this is exactly the kind of thing where any opinion in opposition to this style will feel like the kind of attack that makes a writer want to push back in a "oh yeah? fuck you" kind of way. I.e. even just my writing this opinion may give an author using the style in question the desire to "double down". Though this conundrum is appropriate (ironic?) - the intensely personal nature of language is part of why I love it.
SEARCH FOR “FILM CRIT HULK” FOR SOME EXAMPLES
Descriptive language is how language evolves, and the internet is the first real regional conflict area that Americans have really ever encountered without traveling.
History, you would have just been in your linguistic local, with your own rules, and differences could easily been attributed to outsiders being outsiders. The internet flattens physical distance.
Thus we have a real parallel to the different regions of Italy, where no one came understand each other, or at least the UK, where different cities have extreme pronunciation differences.
The same exists for written language, and it will continue to diverge culturally. The way I look at it is that language isn’t a thing, trapped in amber, but a river we are all wading through. Different people enter at different times, and we all subtly affect the flow.
I distinctly remember thinking “email” was the dumbest sounding word ever. Now I don’t even hear it.
It’s still fine to nitpick, we’re all battling in the descriptive war for correctness. My own personal hobbyhorse is how stupid American quotations syntax is, when learning at graduate school in the UK that you use single quotes and leave the punctuation outside of the quoted sections, which is entirely sensible!
Which, of course, is to donate money to Sama so he can create AGI and be less lonely with his robotic girlfriend, I mean...change the world for the better somehow. /s
Then you can think about automated labs. If things pan out, we can have the same thing in chemistry/bio/physics. Having automated labs definitely seems closer now than 2.5 years ago. Is cost relevant when you can have a lab test formulas 24/7/365? Is cost a blocker when you can have a cure to cancer_type_a? And then _b_c...etc?
Also, remember that costs go down within a few generations. There's no reason to think this will stop.
In that bright AGI future, who does my business serve, like who actually are my actual paying clients? Like, the robots are farming, the robots are driving, the robots are "creating" and robots are "thinking", right? In that awesome future, what paid jobs do us humans have, so my clients can afford my amazing entrepreneurial business that I just bootstrapped with the help of 100s of agents? And how did I get the money to hire those 100s of agents in the first place?
> Is cost a blocker when you can have a cure to cancer_type_a? And then _b_c...etc?
Yes, it very much is. The fact that even known and long discovered solutions like insulin for diabetes management are being sold to people at 9x its actual price should speak to you volumes that while it's great to have cures for X, Y and Z, it's the control over the production and development of these cures that is equally, if not much more important for the cure to actually reach people. In this rosy world of yours, do you think Zuck will give you his LLAMAGI-generated cancer cure out of the goodness of his heart? We are talking about the same dude that helped a couple of genocides and added ads in Whatsapp to squeeze the last cent of the people who are trapped with an app that gets progressively worse and more invasive.
https://www.rand.org/news/press/2024/02/01/index1.html
https://systemicjustice.org/article/facebook-and-genocide-ho...
> Also, remember that costs go down within a few generations. There's no reason to think this will stop.
The destruction of the natural world, the fires all around us, the rise of fascism and nationalism, the wars that are spawning all over the place and the fact that white and blue collar jobs are being automated out while soil erosion and PFAS make our land infertile point to a different future. But yeah, I am simply ecstatic at the possibility that the costs of generating a funny picture Ghibli style with a witty caption could go down by 10 to 30%.
I am seeing problems with formatting that seemed 'solved' already.
I mean, I have seen "the same" model get better and worse already.
clearly somebody is calibrating the stupidity level relative to energy cost and monetary gain
If I (and billions others) can be bothered to learn your damn language so we can all communicate, do us a service and actually use it properly, FFS.
To flaunt:
> display (something) ostentatiously, especially in order to provoke envy or admiration or to show defiance.
To flout:
> openly disregard (a rule, law, or convention).
(I'm also a non-native speaker)
(But then I saw he used the formation - 'Honestly?' which made me think he WAS using LLMs!)