Comment by Tiberium
2 months ago
This is quite a misleading title because this is the raw API cost, but he (obviously) has unlimited usage as an OpenAI employee. Moreover, if you use e.g. the $200 Codex sub, you get about ~$5k-$6k monthly API usage if you spend every week of your usage, if not more, which shows that the raw API cost is not how much it (likely) costs to OpenAI, unless they're subsidizing all this.
He did clarify that it was with fast mode. Without fast mode it'd "only" be $300k in raw API cost, or ~60 $200 Codex subscriptions.
Hey guys, I’m super good at using tokens.
Business: Amazing, that’s great what did you do?
I ran 50 instances and had them all fix the same bugs at the same time and then analyzed the results of all 50 runs to have AI score each of the attempts, then sort them, then compare them to each other in a round robin tournament style double elimination to ensure I got the best result. Then I had AI convert this into a skill, and then ran all 50 attempts again and repeated the process to ensure that I had the absolute best result. It was amazing and I used 1.3 billion tokens!
Business: That is amazing! What did you fix?
A spelling mistake on the About page.
This is the best use of tokens I've ever read. I'm building this skill as we speak to use on our Enterprise Claude account.
Wish me luck on a raise!
Claude, somewhere in this codebase I've mispelled a common word; the word is also a homophone and further, is easily confused with another word that has three r's; please start up a subagent for each file and count the r's and verify how many r's there are's; if there's three, then make sure to review potential homophones and check that I've spelt the correct worrrd incorrectly correct.
Peter Steinberger is of course currently employed by OpenAI and it probably benefits them for him to find ways for their customers to do that.
Sounds like a brilliant promo project!
How is it misleading if this would be the consumer's cost?
Eventually Codex's subscription subsidization will diminish to near-zero, like the rest of the providers.
It's extremely important that people understand how expensive these models currently are. Even $300k in raw API costs is alarming for the output.
> How is it misleading if this would be the consumer's cost?
Because it does not say “equivalent of”, it literally says he spent money that he did not spend
This. If I go up to my boss and say “I spent $10000 but it only cost us $1000” then I spent $1000.
2 replies →
If anything it cost more than the title, because customer costs are wildly subsidized.
So yeah its misleading but in the other direction.
1 reply →
We know how expensive the (Chinese) models are to run, because there are a hundred inference providers selling them cheaply and competitively.
The money going to the American model companies is not going to their hosting costs.
Peter shows the near-term future. Raw API consumer price cost is arbitrary. (The frontier labs can put a 100x markup to cover other operational expenses.) The true cost of inference with same-capability models keeps dropping at dizzying rates, especially at the data-center batch size. (Due to both NVidia hardware and algorithmic changes.) So the developments that Peter can achieve today with internal support from OpenAI will be doable by anyone in a few years without breaking the bank.
But.... why? Like I read his thing on how he spends the tokens [0] and it sounds like satire.
He has agents write shitty code for features other agents think other people want, then has it reviewed by other agents in hopes of catching bugs that the first agent put there, then has some more agents try to find security bugs in the now double-agented code to make it triple-agented and at the end of the day, he spent a shitton of tokens, probably emitted enough carbon to heat our planet by another degree, and has a feature nobody really asked for that might or might not work.
He then has the sense of humor to call this grotesque process "incredibly lean".
What's the point in all of this? What problems is this solving? Who's benefiting?
[0] https://xcancel.com/steipete/status/2055405041843052792
27 replies →
Peter shows shit. What did Peter meaningfully achieve? What additional revenue is he creating? ah yes - shit and more shit on all accounts as it seems.
3 replies →
Even at unlimited budget, there is a crossover where outsourcing thinking to the machine costs more than the machine.
What I mean by this:
1. Intern, analyst, junior, or offshore level coding is cheaper when done by the machine.
// Side note: There is good reason the industry invests in suboptimal output from this set which moves to the "cost" column when using an LLM, but nobody's accounting for that.
2. For the interns, analysts, junior, or offshoring to do the right thing costs a multiple of the coding effort: the PdM/PjM stuff of course, but also the Stakeholder, Product Owner, Architect, Principal Engineer, QA, and SRE stuff.
3. If you are not a principal or staff engineer level engineer, you are likely unqualified to catch and fix the errors LLMs make across engineering, much less these other PDLC (product development lifecycle, which includes SDLC and SRE) loop.
4. For LLM output to be useful, your 'harness' has to incorporate all of that as well, which because it's so much harder than transliterating spec-to-code, balloons tokens exponentially.
5. Today it is faster, more efficient, and costs less, to work with LLMs "XP" (eXtreme Programming) style, pairing with the LLM actively co-creating and co-reviewing, steering for more effective turns.
So, your options are:
- ship garbage while costing less than a median first world SWE
- pair with the LLM actively for the benefits of XP
- add enough harness and steering the LLM costs more than SWEs, and still needs a human loop “move fast and break things to find out what's broken” style
I would expect that within a couple years, these other disciplines can be baked in enough the machine costs less for everything but surprises.
> I would expect that within a couple years, these other disciplines can be baked in enough the machine costs less for everything but surprises.
They already are. I’m successfully using frameworks like bmad to deliver complex apps at that level. My job is to manager the see, as, ux, sre processes and catch errors.
I spend more time refinding prd , epics and stories than I do elbows deep in code.
If I don’t like the output of a story I nuke it change the story and have the flanker try again. I’m using the open source glm, kimi, deepseek models. I expect the full pipeline to be good enough by the end of the year.
> I spend more time refinding prd , epics and stories than I do elbows deep in code.
And do you enjoy this more than writing code? I used to look forward to writing code, solving these little optimization puzzles, learning, and staying sharp. Working with agents is dreadful in comparison. They lie, rarely learn, and I feel like a proctor.
Sure, you sometimes get to see something amazing, but usually I am just very annoyed by their performance and ever-changing but never-ending billing issues. First, with Claude Code, now with Codex, which was fine for a minute, but now I am out of tokens for the majority of time. (I don't have the income for those Pro INTx plans.)
1 reply →
I think its less misleading this way because every other reader would have to pay $1.3M to emulate his workflow for a similar size project. His discounted internal costs are relevent only to openai.
I did mention that you could use ~60 $200 Codex accounts to emulate his workflow without /fast, or 2.5x that if you used /fast. Not $1.3M
"The tokens are how anthropic makes profit" vs "It's not actually worth that amount of usage"
ya'll cant have it both ways; either it's really worth the cost or it's a bunch of token burn with no smoke.
> unless they're subsidizing all this
They literally are. (If by "all this" you mean the subscription future bait-and-switch plans.)
But even going with the $5k - $6k monthly usage on a $200 codex subscription even going over their limits is also unrealistic in the long term and that is just ONE person.
Lets say I was at the casino and was spending a lot on casino chips but I also happen to work at the casino. I'm not really losing money whether if I win / lose since I'm using the houses money and there's little risk involved on every dice roll or press of the button. The risk is far higher if I don't have that level of access and continue to spend the same amount of money on lots of tokens (or casino chips, spins or button presses.)
The same is true here with these agents. Some companies will realize that they can no longer afford to spend millions a month on tokens or even startups spending $5k - $6k per person per month on tokens.
I can only see local efficient models making sense on recovering from this unnecessary spending or even light gambling on tokens.