Hacker News new | ask | show | jobs
by thomasahle 31 days ago
He used 600B tokens in 30 days.

I use more than 150B/month with just 15 codex accounts.

60 accounts is "just" $12,000/month. So Peter could "save" 100x by using monthly accounts.

Of course, he doesn't have to, as he works at OpenAI now.

3 comments

Sounds like a healthy industry, selling tokens at 1000x below cost.
API pricing isn’t cost, we don’t know what cost is.
I would bet money Anthropic and OpenAI are actually profitable on inference. The problem is they have to spend large sums of money to train models that are essentially worthless after a few months.
Dario explicitly stated this in an interview.

They make more money from inference than they do training the model, but then the next model gets so much more expensive to train so their annual figures have been in the red.

So, it's like if they were a pharma company that was barely profitable if you didn't take into account R&D costs?
A large part of the GPT-5.x model iteration has been about making training more affordable and token efficient.
It's to build a moat, of course!

Narrator: there was no moat

This performative concern over token costs and subsidisation comes from either ignorance or some latent ideology signalling.
One could say "that's a great point, we should take more direct ideological action to address this issue!", but expounding upon the finer details would likely get one banned here.
What I truly don't understand, as a daily heavy Opus 4.7 user, is how you can coherently prompt 15 different parallel conversations at the same time.

For me it's not even a "what the hell are you working on" so much as complete inability to understand how you can keep so many different processes working on distinct tasks. It simply doesn't map on to how I use these tools.

I spend most of my day writing extremely detailed prompts and that's how I'm able to get the sort of excellent results that confound skeptics. But I have to be honest with you: I don't think I can write (or think) fast enough to do two of these at a time, much less 15.

I definitely could not review what they are generating with any degree of confidence.

I'm really hoping you can explain what the heck your usage pattern actually looks like, because reading this makes me feel like I'm missing something.

I'm trying to recreate all the commercial EDA stack in open source. (RTL simulators, synthesis, formal proof tools, etc.)

Building compilers has a _lot_ of parallel tasks agents can work on.

Wish me luck..

Good luck!
Yeah good luck with that. I find SystemVerilog is probably the thing that AI is worst at, presumably because there's not that much training data out there, and pretty much everything about the commercial tools is paywalled.
those costs are not just tokens used for prompting . costs include agent loops, etc
What do you do with all those accounts?
Probably trying to fix their broken personal website with the half of the links there not working at all.
Is my website broken?
Ask your 15 codex accounts agents, surely they will help you with that.
Your website seemed fine to me I didn't try every link though