when I still had access to flat rate seats at work (I think equivalent to the consumer 200/mo) I was at estimated ~8000-ish/month api costs at 100% usage, which is 400/workday. This with maxing out the 50% fable usage and using opus for the rest, usually 1-3 concurrent sessions.
Having moved to api billing (where my spend is not allowed to be nowhere near that high), I think that estimate was conservative and I wouldn't be surprised if it was more like 500/workday equivalent.
I am glad I was able to do a lot of experimenting and learning while we were on flat rate, and I am surprised a lot of coworkers did not take advantage while they could.
I also have the 20$ pro plan, I'm not reaching the limits of it... That said, I'm a one man company so I'm not just emptying a kanban board, I use Claude a lot as a thinking buddy, and my deliberations are the thing that limit the speed.
It's kinda borderline. My dev workflow in my full-time job (Opus 5.5, 5 concurrent worktrees with human checking results and giving instructions) spend exactly $200/day on average. I imagine it could be more friendly if it's $300, as otherwise I might have to move some of my AI usage to weekends, which implicitly increases working hours :[
In my experience Claude Max x5 is the very reasonable amount of tokens for a work week.
But yeah, API pricing is crazy. Note that Apple might have negotiated a discount with Anthropic but is actually not allowed to share this discount details with its employees. So they communicate the budget based on public pricing.
$52K per engineer per year extra cost to the company. I think at an Apple scale that level of ROI should be easy. I wonder if they were not seeing the ROI they wanted? As an indie hacker AI has been amazing for me and has freed me up to do so much more.
Or, get this - figuring out what customers really want and need is super hard and churning out code doesn't solve the truly hard problems of software development.
I don't see any indication that LLMs help with requirements gathering or user studies, especially live ones.
Yes, it's a significant expense even on top of an Apple salary.
Any company, not just Apple, should be tracking whether the extra spending is worth it. There must be a point of diminishing returns where additional AI usage stops increasing productivity at a comparable rate. That point should be your target. Different teams can have vastly different thresholds.
Companies also can't justify ever-increasing costs just because it increases productivity. It's business. You care about profits, not raw productivity. And Apple doesn't seem to be having a particularly great time these days.
They are looking to lower it further to $100 [0]. This is definitely concerning for frontier labs since Accenture also caps their employees' usage to $29 a month [1].
$100 day should be more than enough. To exceed that you have to coming up with creative ways to waste it. You couldn’t hit that in normal development using AI all day.
For the edge cases that might need more like running a security audit, it would make more sense to submit that as a project proposal and have the expense approved.
I don’t agree. You should be empowering your engineers. A higher budget allows them to:
- freely experiment and understand the limitations of AI
- create things outside of what they are officially expected to
I’m not advocating for people who burn tokens for the sake of it. But I do think a single engineer has more potential for impact today than ever before and a high budget unlocks that.
It also isn’t _that_ much more when a fully loaded mid-level engineer is costing you $300k+
I find it hard to use that much in a day. Especially if you delegate to a cheaper model for anything context hungry like reading large amounts of data etc
token-maxxing aside, what's the business value for something like this? Slide-deck generators have been around for a long time already. And at least for me, the LLM-authored decks I've created have always needed tweaking / editing to really make the content correct and punchy.
> what's the business value for something like this?
We already generated slide decks from a template. The decks generated were consistently missing data, incorrectly formatted, and required many manual edits.
This led to hiring contractors to fix up the decks before being sent internally before being presented to clients.
So, thousands of contractor hours every year going towards fixing up decks in additional to the engineering effort from ~5 bugs each month due to bugs in the deck or feature gaps.
> Then they use tool a few times and ultimately forget about it or throw it away.
This is a running production service
> More time gets spent tweaking/iterating with the LLM than it would take to make the decks by hand.
I think you misunderstood. The LLM produced deterministic deck templates in code that is then templated from a payload and the deck dynamically changes depending on what is in the payload. The user can then further modify the deck as they please since the deck is just code. Then they can export it to a PPTX and make any further adjustment in their favorite tool since it emits a "normal" PPTX.
Techie types seem particularly susceptible to spending massive effort re-inventing things that have already existed for a while.
See also, Matt Levine’s recurring joke about the crypto community constantly re-inventing/re-discovering the need for financial regulations that the larger finance industry have had for a while.
> $100 day should be more than enough. To exceed that you have to coming up with creative ways to waste it.
You do not need much creativity for that. Just forget to put build artifacts or node_modules in your agent's ignore list and it will happily burn through the cap in thirty minutes
Yes. The frontier models are good, indeed, but they are just far too expensive to use much. Recently I was showing my executive leader that we could simply run a certain set of tasks on the top frontier model and we'd be good to go. Oh but it would cost over 15M/year. We're a relatively small company, that's over two magnitudes more than what makes sense for us. So, never going to happen.
What remains then is an optimization exercise to bring the cost down. For months now we've been optimizing usage so we can run nearly all tasks on very cheap models to good success. It's fun to optimize and keeps the CFO from having a heart attack.
But not good news for the AI companies relying in a firehose of cash coming in from their frontier models.
It's kinda like commuting in the autobahn with a Porsche.
Some agentic madness is like doing donuts in the parking lot drinking beer impressing the local baddies.
So, so much fun. But at the end of the day we used to get so much more done with much, much less computer power, energy, mental drain.
It's always more software tooling, more moving parts, more apps, more more more. Fun. Looks neat. Used sooo much energy, I feel so drained. It was fun.
But at the end of the day, for the long run what I want in my life is to win at rally on a dependable machine with a nice navigator I'm friends with. Or have a shed full of tractors making money for my family on the farm everyday.
LLMs are fun. But I want to use my mind to write great software that adds value in as little lines of code as possible designing systems that people can understand readily and maintain well 2 or 3 handovers from me.
Why is it "concerning"? Personally, I get a lot of stuff done on the $200/month subs. +$100/day per employee seems like a fairly significant amount of money and if a lot of companies do that (even a lot less than that) that seems entirely fine for frontier labs, given all the other uses as well.
That is not to say that you can't spend a lot more. And I would be surprised if Apple did not allow for that too. Note that the text says "for some of its own engineers", which could amount to a total nothing-burger, given the overall vagueness.
Depends on the sub you're using, but the anthropic one is giving you ~$12k/month [0], or $400/day. So you're already working at 4x the $100/day you consider significant.
I’m not sure if the /usage in cc is directly comparable to this? It gives a $ value which I assume is API equivalent.
Most of the programming work I get opus5.5 to do, sometimes with fable coordinating. When I’m building a large feature it will often run over $500 and occasionally on longer sessions over 1k. Everyone else seems to think $200 should get by?
Just checked codexbar. I spend $1,000/mo in API credits across two $20/mo plans (claude and codex) just for side projects. I can't imagine having a API budget in this range for a day job, I'd run out every day
Do you work at a company that requires human code review for every change? Even if I tried to spend that much my work would just get logjammed in code review, so I intentionally don't generate output faster than I can get reviewers to read it.
Finally companies are realizing that subsidizing unlimited AI for your employees is basically just paying for the tools for them to work 2hrs a day while you pay them for 8.
Listen, I dont care if employees scam their bosses by offshoring their work to AI, but its absolutely the case that 90% of people irresponsibly use AI to be lazy, get by doing less work.
Unfortunately, as Ive personally witnessed when you do this and take away the cheating tools engineers just start copying and pasting IP into their own AI subs where you have less garuntees about ZDR. Employees will gladly spend an extra $200 a month so they can work 4hrs less a day and just get by.
Only 10%, maybe less, are actual super contributors with AI. The rest are literally getting a form of diabetes of the brain and the exercise required to reverse that disease is quite uncomfortable, especially when you've been hands off using a cheating tool for a year straight.
I wont argue with you there, I dont really care about the companies at all
Im just saying employers will notice that its basically being used as a cheating tool by a large subset of users and they 'll stop subsidizing that behavior.
What Im really trying to say is its not a producitivity enhancer for an enterprise, maybe for a small percentage of employees, so I dont really see the value in it as an enterprise wide investment.
200 daily means that you use 1 session in a turn-based setup. Hence likely you still know what is going on in the code.
Once you have multiple agents working in parallel, hands off the wheel becomes very attractive. And at that exactly point you lose completely touch of what the heck is going on in your ex-code.
None of the financials make sense right now because everyone is racing to AGI under the assumption that whoever wins will (probably literally) rule the world
Giving people an unlimited amount of anything means they have zero incentive to optimize using that thing. It's quite common that we complain here about Electron apps using 1GB for something as simple as a REST client. OTOH, many here keep calling for unlimited token budgets. I feel like capping will cause people to find better workflows, which we really need to in order to reduce (among others) the ecological impact of AI.
86 comments
[ 0.22 ms ] story [ 53.3 ms ] threadHaving moved to api billing (where my spend is not allowed to be nowhere near that high), I think that estimate was conservative and I wouldn't be surprised if it was more like 500/workday equivalent.
I am glad I was able to do a lot of experimenting and learning while we were on flat rate, and I am surprised a lot of coworkers did not take advantage while they could.
But yeah, API pricing is crazy. Note that Apple might have negotiated a discount with Anthropic but is actually not allowed to share this discount details with its employees. So they communicate the budget based on public pricing.
I appreciate the boundless optimism.
Or, get this - figuring out what customers really want and need is super hard and churning out code doesn't solve the truly hard problems of software development.
I don't see any indication that LLMs help with requirements gathering or user studies, especially live ones.
You cut it when you have evidence that it was wasteful.
Any company, not just Apple, should be tracking whether the extra spending is worth it. There must be a point of diminishing returns where additional AI usage stops increasing productivity at a comparable rate. That point should be your target. Different teams can have vastly different thresholds.
Companies also can't justify ever-increasing costs just because it increases productivity. It's business. You care about profits, not raw productivity. And Apple doesn't seem to be having a particularly great time these days.
They vibe coded some slopware mobile app on a $200 Claude subscription and roleplay as that being productive.
In reality those companies are measuring what the extra cost brought in as a return, and are likely deciding that it was not worth it.
I fully expect this to become the norm, figuring out ways to extract usefulness out of LLMs while spending as little as possible.
[0] https://finance.yahoo.com/technology/ai/articles/apple-cuts-...
[1] https://finance.yahoo.com/technology/ai/articles/accenture-j...
For the edge cases that might need more like running a security audit, it would make more sense to submit that as a project proposal and have the expense approved.
It's easy to spend a lot when writing an entire service, large features, or doing research loops.
- freely experiment and understand the limitations of AI
- create things outside of what they are officially expected to
I’m not advocating for people who burn tokens for the sake of it. But I do think a single engineer has more potential for impact today than ever before and a high budget unlocks that.
It also isn’t _that_ much more when a fully loaded mid-level engineer is costing you $300k+
- Convert PPTX to TypeScript
- Allow editing of that TypeScript via an LLM interface
- Convert TypeScript to PPTX
The intent was to create slide deck templates via code. The first step is to author the templates which the LLM did for me perfectly.
It is quite hard to convert an arbitrary PPTX into code and back.
So I had it write a harness, gave them problem to Opus, and a couple thousand dollars later my team had a solution to a problem plaguing us for years.
We already generated slide decks from a template. The decks generated were consistently missing data, incorrectly formatted, and required many manual edits.
This led to hiring contractors to fix up the decks before being sent internally before being presented to clients.
So, thousands of contractor hours every year going towards fixing up decks in additional to the engineering effort from ~5 bugs each month due to bugs in the deck or feature gaps.
> Then they use tool a few times and ultimately forget about it or throw it away.
This is a running production service
> More time gets spent tweaking/iterating with the LLM than it would take to make the decks by hand.
I think you misunderstood. The LLM produced deterministic deck templates in code that is then templated from a payload and the deck dynamically changes depending on what is in the payload. The user can then further modify the deck as they please since the deck is just code. Then they can export it to a PPTX and make any further adjustment in their favorite tool since it emits a "normal" PPTX.
See also, Matt Levine’s recurring joke about the crypto community constantly re-inventing/re-discovering the need for financial regulations that the larger finance industry have had for a while.
- Emits normal PPTX
- Can import any PPTX
- Representation can easily be edited by an LLM
- Handles most PPTX features -- rich text, transforms, charts/tables, images
- Can be templated
Things I tried:
- python-pptx, PptxGenJS
- Marp
- Google Slides + Google Sheets + Google AppScript
You do not need much creativity for that. Just forget to put build artifacts or node_modules in your agent's ignore list and it will happily burn through the cap in thirty minutes
Yes. The frontier models are good, indeed, but they are just far too expensive to use much. Recently I was showing my executive leader that we could simply run a certain set of tasks on the top frontier model and we'd be good to go. Oh but it would cost over 15M/year. We're a relatively small company, that's over two magnitudes more than what makes sense for us. So, never going to happen.
What remains then is an optimization exercise to bring the cost down. For months now we've been optimizing usage so we can run nearly all tasks on very cheap models to good success. It's fun to optimize and keeps the CFO from having a heart attack.
But not good news for the AI companies relying in a firehose of cash coming in from their frontier models.
It's kinda like commuting in the autobahn with a Porsche.
Some agentic madness is like doing donuts in the parking lot drinking beer impressing the local baddies.
So, so much fun. But at the end of the day we used to get so much more done with much, much less computer power, energy, mental drain.
It's always more software tooling, more moving parts, more apps, more more more. Fun. Looks neat. Used sooo much energy, I feel so drained. It was fun.
But at the end of the day, for the long run what I want in my life is to win at rally on a dependable machine with a nice navigator I'm friends with. Or have a shed full of tractors making money for my family on the farm everyday.
LLMs are fun. But I want to use my mind to write great software that adds value in as little lines of code as possible designing systems that people can understand readily and maintain well 2 or 3 handovers from me.
That is not to say that you can't spend a lot more. And I would be surprised if Apple did not allow for that too. Note that the text says "for some of its own engineers", which could amount to a total nothing-burger, given the overall vagueness.
[0]: https://newsletter.semianalysis.com/p/anthropic-subscription...
Most of the programming work I get opus5.5 to do, sometimes with fable coordinating. When I’m building a large feature it will often run over $500 and occasionally on longer sessions over 1k. Everyone else seems to think $200 should get by?
I’m looking for a point of reference here.
Listen, I dont care if employees scam their bosses by offshoring their work to AI, but its absolutely the case that 90% of people irresponsibly use AI to be lazy, get by doing less work.
Unfortunately, as Ive personally witnessed when you do this and take away the cheating tools engineers just start copying and pasting IP into their own AI subs where you have less garuntees about ZDR. Employees will gladly spend an extra $200 a month so they can work 4hrs less a day and just get by.
Only 10%, maybe less, are actual super contributors with AI. The rest are literally getting a form of diabetes of the brain and the exercise required to reverse that disease is quite uncomfortable, especially when you've been hands off using a cheating tool for a year straight.
Im just saying employers will notice that its basically being used as a cheating tool by a large subset of users and they 'll stop subsidizing that behavior.
What Im really trying to say is its not a producitivity enhancer for an enterprise, maybe for a small percentage of employees, so I dont really see the value in it as an enterprise wide investment.
Once you have multiple agents working in parallel, hands off the wheel becomes very attractive. And at that exactly point you lose completely touch of what the heck is going on in your ex-code.
I have been able to soak up a weeks worth of tokens in 6 hours with astra ultra if I'm not very explicit and bounded in my requests.
But we already know they have their own models, ship their own IDE and have released many programming languages.
It seems clear that they would be negligent if they weren’t training on all this employee token usage that they are paying for