[dpcan]

I cannot figure out what people are doing to spend all this money.

I have used a $60 per month Cursor plan on auto, and have never come close to using up my included usage, and I probably have it planning and coding and working for me all through the evenings 4 nights a week.

What on earth are people doing differently that it's costing them so much?

Maybe enabling on-demand usage or other paid models, or on higher modes? What are you doing that requires this? The output from Auto for me is crazy good for the tasks I'm working on, and have yet to run into an issue where it couldn't perform at a high enough level.

We have been interviewing people at work to join our team and they tell us they use $2K per month in tokens with their current employers.... I can't even fathom what's going on here where that would be happening.

[shepherdjerred]

If you can give your agent broad access and an effective feedback loop, you just need to steer it and do a final check on outputs.

As an example I might have an agent with access to a browser, logs, metrics, GitHub& CI logs etc. and ask it to implement a new feature.

In Slack I have a few bug reports so I spin up a few more agents. A PM needs a UI tweak so I spin up an agent. You can imagine that a lot of work a dev does isn’t necessarily that complicated and I just need to be there to review the final PR and leave comments as if it were a colleagues (and then my agent goes back, fixes the comments, requests a new review…)

While that’s happening I might be using my actual attention for a meaty feature, design doc, data analysis, etc.

I spend $300/mo for personal use, and a couple thousand at work. Agents can be really transformative and well worth the cost.

Would my company rather pay a few thousand per month, or a several hundred thousand per year for an extra fully loaded engineer? At this point it is _at least_ a 2x multiplier for myself

[timr]

> Would my company rather pay a few thousand per month, or a several hundred thousand per year for an extra fully loaded engineer?

I don't know, and neither do you. You haven't cited a return for all of this expense, so the math can't be done [1].

In the old times, we couldn't hire enough intelligence to do everything we wanted to do, at any price. So we prioritized what we really wanted, and the rest didn't get done. This was largely fine, or at least, we didn't have the time to consider paths that were left unexplored [2].

Now we have a way to spend essentially unbounded amounts of money for intelligence on demand [3]. So we don't have to prioritize anymore -- we can set the priority bar for execution arbitrarily low, so long as we're willing to spend money to do it [4]. So what is the right amount to spend per unit of intelligence? What is the value produced per unit of expense? This is an entirely new problem. No other business efficiency technology, save perhaps electricity or steam, has had this kind of uncapped upper bound.

Nobody knows the answer, but I strongly suspect that the answer is closer to "far less than everyone thinks it is right now, when everyone is trapped in a prisoner's dillema". After all, you don't want to be the company that gets caught out by your competitors who are willing to spend the extra marginal dollar [5].

I don't have a solution either, but I think it's worth calling out that the fallacy is an elephant-sized, power-guzzling problem, and it isn't being helped by the current industry pricing models, which are encouraging token-guzzling. If you had to press me, I think we'll eventually come to our senses and get back to the same basic approach companies use to allocate employee spend today -- the budget allocation for "engineering" is periodically re-bid according to net observed product velocity, with some multiplier applied for "increased efficiency of robot-augmented human intelligence" [6].

[1] and let's be brutally honest: you didn't cite a return, because you have no idea how to measure it.

[2] Maybe the business-changing transformation was one commit away all along! Horrors!

[3] Worse, there are price tiers for "quality" of intelligence, which also cannot be defined or quantified.

[4] Not really, of course -- there's still a limit in the number of bots that a human can manage, the number of GPUs in the world, etc. but for the sake of argument, it's an unlimited upper bound on intelligence.

[5] Again: horrors!

[6] It's obvious that the multiplier is not "infinity".

[isubkhankulov]

Claude enterprise plans are 30-40x more expensive vs the consumer plans.

I used to spend $200/mth on the Max plan at a small startup. Now spending single digit thousands on Claude enterprise with the same usage levels.

Anthropic is subsidizing consumer usage, and also charging a nice margin for enterprises for zero data retention (ZDR)

[cubano]

> and they tell us they use $2K per month in tokens with their current employers...

perhaps they are simply trying to impress you with their mad prompting skills and like, what self-respecting engineer would be caught dead using less then $2k/month?

giving the context of your interaction with those people, it probably is the simplest answer to your rather baffling question. for the life of me the idea of using $2k/month doesn't even seem possible unless your telling it to waste credits.

[galaxyLogic]

Sounds about right. Those people seem to want to create the impression that they are "AI Power Users". That gives them more power inside the organization. People come to them to ask for advice. Also if their output is not good they can claim that is because the AI budget didn't allow them to do more.

[jmkni]

Totally agree, but then a lot of the same people will be talking about all of the custom instructions/rules/skills/features etc they have set up, so that's eating up a lot of the context window before you even start

When I do use AI, it's just the pure tool itself, and the context is the exact code I'm working with (because I'm trying to see if it can help me solve a specific problem), and I understand the rest of the codebase well enough to know if it's giving me good answers or bad ones

[rjh29]

A few things imo, 1) not prompting precisely enough (narrowing scope) means your agent will scan your entire code-base and sometimes get stuck looking at things repeatedly. 2) not checking the output is usually fine but sometimes it produces junk because it doesn't understand, and you cannot prompt your way out of it without reading the code and figuring out the problem. If you leave it on auto it will burn tokens.

Plenty of low level things can trip agents up, too. I just had one inexplicably refuse to read an error about a function needing a bool return value - trying about 10 variations of the same thing before I interrupted it. Skills probably cause issues too, it loves to for example read the source code of libraries I'm using if I give it permission. That's a rabbit hole.

[sbochins]

To be honest I’m not doing too much. I’m just on one of the $200 plans, but always hit limits. I only use the best models and mostly use it for various software projects I always wanted to build, but didn’t have time for. I just closely monitor the usage caps and have something running on a Ralph loop most of the time, unless I get near the cap. The post here is more about how I’d start a self-funded software company, if I wasn’t already working full time.