Comment by e1g
20 hours ago
Yes, my CC usage is regularly $50-$100 per day, so their Max plan is absolutely great value that I don’t expect to last.
20 hours ago
Yes, my CC usage is regularly $50-$100 per day, so their Max plan is absolutely great value that I don’t expect to last.
Pretty easy to hit $100 an hour using Opus on API credits. The model providers are heavily subsidized, the datacenters appear to be too. If you look at the Coreweave stuff and the private datacenters it starts looking like the telecom bubble. Even Meta is looking to finance datacenter expansion - https://www.reuters.com/business/meta-seeks-29-billion-priva...
The reason they are talking about building new nuclear power plants in the US isn't just for a few training runs, its for inference. At scale the AI tools are going to be extremely expensive.
Also note China produces twice as much electricity as the United States. Software development and agent demand is going to be competitive across industries. You may think, oh I can just use a few hours of this a day and I got a week of work done (happens to me some days), but you are going to end up needing to match what your competitors are doing - not what you got comfortable with. This is the recurring trap of new technology (no capitalism required.)
There is a danger to independent developers becoming reliant on models. $100-$200 is a customer acquisition cost giveaway. The state of the art models probably will end up costing hourly what a human developer costs. There is also the speed and batching part. How willing is the developer to, for example, get 50% off but maybe wait twice as long for the output. Hopefully the good dev models end up only costing $1000-$2000 a month in a year. At least that will be more accessible.
Somewhere in the future these good models will run on device and just cost the price of your hardware. Will it be the AGI models? We will find out.
I wonder how this comment will age, will look back at it in 5 or 10 years.
Your excellent comments make me grateful that I am retired and just work part time on my own research and learning. I believe you when you say professional developers will need large inference compute budgets.
Probably because I am an old man, but I don’t personally vibe with full time AI assistant use, rather I will use the best models available for brief periods on specific problems.
Ironically, when I do use the best models available to me it is almost always to work on making weaker and smaller models running on Ollama more effective for my interests.
BTW, I have used neural network tech in production since 1985, and I am thrilled by the rate of progress, but worry about such externalities as energy use, environmental factors, and hurting the job market for many young people.
I've been around for a while (not quite retirement age) and this time is the closest to the new feeling I had using the internet and web in the early days. There are simultaneously infinite possibilities but also great uncertainty what pathways will be taken and how things will end up.
There are a lot of parts in the near term to dislike here, especially the consequences for privacy, adtech, energy use. I do have concerns that the greatest pitfalls in the short terms are being ignored while other uncertainties are being exaggerated. (I've been warning on deep learning model use for recommendation engines for years, and only a sliver of people seem to have picked up on that one, for example.)
On the other hand, if good enough models can run locally, humans can end up with a lot more autonomy and choice with their software and operating systems than they have today. The most powerful models might run on supercomputers and just be solving the really big science problems. There is a lot of fantastic software out there that does not improve by throwing infinite resources at it.
Another consideration is while the big tech firms are spending (what will likely approach) hundreds of billions of dollars in a race to "AGI", what matters to those same companies even more than winning is making sure that the winner isn't a winner takes all. In that case, hopefully the outcome looks more like open source.
The SOTA models will always run in data centers, because they have 5x or more VRAM and 10-100x the compute allowance. Plus, they can make good use of scaling w/ batch inference which is a huge power savings, and which a single developer machine doesn’t make full use of.
> Pretty easy to hit $100 an hour
I don’t see how that can be true, but if it is…
Either you, or I are definitely use Claude Code incorrectly.
It’s definitely easy with an API key I hit 200$ in an evening. I didn’t think that could be possible. Horrifying
3 replies →
This is around what what Cursor was costing me with Claude 4 Opus before I switched to Claude Code. Sonnet works fine for some things, but for some projects it spews unusable garbage, unless the specification is so detailed that it's almost the implementation already.
2 replies →
Why “no capitalism required”? Competition of this kind is only possible with capitalism.
Have you been human before? competition for resources and status is an instinctive trait.
It rears its head regardless of what sociopolitical environment you place us in.
You’re either competing to offer better products or services to customers…or you’re competing for your position in the breadline or politburo via black markets.
Even in the Soviet Union there were multiple design bureaus competing for designs of things like aircraft. Tupolec, Ilyushin, Sukhoi, Mikoyan-Gyurevich (MiG), Yakolev, Mil. There were quite a lot. Several (not all, they had their specialisations) provided designs when a requirement was raised. Not too different from the US yet not capitalist.
Unfortunately it's called war and it appears to be part of human nature.
Not really, it's possible with any market economy, even a hypothetical socialist one (that is, one where all market actors are worker-owned co-ops).
And, since there is no global super-state, the world economy is a market economy, so even if every state were a state-owned planned economy, North Korea style, still there would exist this type of competition between states.
6 replies →
Can you give me an idea of how much interaction would be $50-$100 per day? Like are you pretty constantly in a back and forth with CC? And if you wouldn’t mind, any chance you can give me an idea of productivity gains pre/post LLM?
Yes, a lot of usage, I’d guess top 10% among my peers. I do 6-10hrs of constant iterating across mid-size codebases of 750k tokens. CC is set to use Opus by default, which further drives up costs.
Estimating productivity gains is a flame war I don’t want to start, but as a signal: if the CC Max plan goes up 10x in price, I’m still keeping my subscription.
I maintain top-tier subscription to every frontier service (~$1k/mo) and throughout the week spend multiple hours with each of Cursor, Amp, Augment, Windsurf, Codex CLI, Gemini CLI, but keep on defaulting to Claude Code.
I am curious what kind of code development you are doing with so many subscriptions?
Are you doing front end backend full stack or model development itself?
Are you destilling models for training your own?
I have never heard someone using so much subscription?
Is this for your full time job or startup?
Why not use qwen or deep seek and host it yourself?
I am impressed with what you are doing.
9 replies →
Mostly to save money (I am retired) I mostly use Gemini APIs. I used to also use good open weight models on groq.com, but life is simpler just using Gemini.
Ultimately, my not using the best tools for my personal research projects has zero effect on the world but I am still very curious what elite developers with the best tools can accomplish, and what capability I am ‘leaving on the table.’
Thank you for your perspective. I’ve been staring at Claude Code for a bit and I think I will just pull the trigger.
1 reply →
I am curious what kind of development you’re doing and where your projects fall on the fast iteration<->correctness curve (no judgment). I’ve used CC Pro for a few weeks now and I will keep it, it’s fantastically useful for some things, but it has wasted more of my time than it saved when I’ve experimented with giving it harder tasks.
1 reply →
Re productivity gains, CC allows me to code during my commute time. Even on a crowded bus/train I can get real work done just with my phone.
How do you use Claude Code via your phone?
2 replies →
What's your workflow if I may ask? I've been interested in the idea as well.
1 reply →
Unless you're getting paid for your commute, you're just giving your employer free productivity. I would recommend doing literally anything else with that time. Read a book, maybe.
2 replies →
you can easily reach 50$ per day. by force switching model to opus /model opus it will continue to use opus eventhough there is a warning about approaching limit.
i found opus is significantly more capable in coding than sonnet, especcially for the task that is poorly defined, thinking mode can fulfill alot of missing detail and you just need to edit a little before let it code.
wow. haven't tried Opus but Sonnet 4 is already damn good.
Shhh don’t say that. I love Max. I don’t want it to go anywhere.
Is there a cheap version for hobbyists? Or what’s the best thing for hobbyists to use, just cut and paste?
Claude Code with a Claude subscription is the cheap version for current SOTA.
"Agentic" workflows burn through tokens like there's no tomorrow, and the new Opus model is so expensive per-token that the Max plan pays itself back in one or two days of moderate usage. When people reports their Claude Code sessions costing $100+ per day, I read that as the API price equivalent - it makes no sense to actually "pay as you go" with Claude right now.
This is arguably the cheapest option available on the market right now in terms of results per dollar, but only if you can afford the subscription itself. There's also time/value component here: on Max x5, it's quite easy to hit the usage limits of Opus (fortunately the limit is per 5 hours or so); Max x20 is only twice the price of Max x5 but gives you 4x more Opus; better model = less time spent fighting with and cleaning up after the AI. It's expensive to be poor, unfortunately.
>less time spent fighting with and cleaning up after the AI.
I've yet to use anything but copilot in vscode, which is 1/2 the time helpful, and 1/2 wasting my time. For me it's almost break-even, if I don't count the frustration it causes.
I've been reading all these AI-related comment sections and none of it is convincing me there is really anything better out there. AI seems like break-even at best, but usually it's just "fighting with and cleaning up after the AI", and I'm really not interested in doing any of that. I was a lot happier when I wasn't constantly being shown bad code that I need to read and decide about, when I'm perfectly capable of writing the code myself without the hasle of AI getting in my way.
AI burnout is probably already a thing, and I'm close to that point already. I do not have hope that it will get much better than it is, as the core of the tech is essentially just a guessing game.
1 reply →
cursor on a $20/month plan (if you burn thru the free credits) or gemini-cli (free) are 2 great ways to try out this kinda stuff for a hobbyist. you can throw in v0 too, $5/month free credits. susana’s free tier can give you a db as well.
Claude Code pro is ~$20USD/ month and is nearly enough for someone like me who can’t use it at work and is just playing around with it after work. I’m loving it.
If you are a hobbyist, just use Google’s gemini-cli (currently free!) on a half dozen projects to get experience.
I've been enjoying Zed lately
Zed is fantastic. Just dipping my toes in agentic AI, but I was able to fix a failing test I spent maybe 15 minutes trying to untangle in a couple minutes with Zed. (It did proceed to break other tests in that file though, but I quickly reverted that.)
It is also BYOA or you can buy a subscription from Zed themselves and help them out. I currently use it with my free Copilot+ subscription (GitHub hands it out to pretty much any free/open source dev).
Cursor at 20$/M is pretty great