[rstuart4133]

> Humans are very expensive, so the equation almost always falls against them.

You underestimate what these models cost. Uber's budget is $1,500/dev/month. I gather that was put in place because the dev's were going through $6,000/dev/month, which Uber decided could not be cost justified.

Fable costs at least twice as much, or $12,000/dev/month.

Fable can apparently work for hours without supervision, which means a skilled engineer can now have it working on many tasks concurrently. I would not be at all surprised if they can put a nought or two on that number. If you do that, you are well out of "what a human costs" territory.

[swiftcoder]

> You underestimate what these models cost. Uber's budget is $1,500/dev/month

$1,500/month needs to be contextualised against the fully-loaded cost of a software engineer. Uber's average TC for a US-based software engineer is around $350k, the fully-loaded cost is going to be in the $450k-$500k range. So we're talking around $38k/month for a software engineer.

$1,500/month isn't even a drop in the bucket. If LLM use lets them shave just one person off a team, that pays for tokens for the next 25 engineers.

[amarant]

Not to argue myself out of a job, but I cost around $20k/month, all costs considered(taxes, social fees, PTO, healthcare, benefits). If my efficiency is tripled(which it absolutely is, even before fable) for a mere 6k/month(in reality, 1k is more than enough though), that's ~10x ROI.

I kinda get why execs are excited

[popcorncowboy]

> If my efficiency is tripled

Our 401ks turn on this actually being true. Otherwise pop.

[nearbuy]

These numbers don't mean anything without a denominator. You could burn $10 million/month of tokens if you want. We want to know how the cost per unit of useful output compares to a human. Does $6000 of usage buy you a man-month of work? Less? More?

[winwang]

Minor note, 2x $/tok is not 2x cost. Personally, I see Fable being significantly more token-efficient than Opus 4.8. Then, there's also the compounding costs of quality.

[squidbeak]

On top of which, as the article mentions, it delegates simpler tasks to cheaper models.

[ramesh31]

> I would not be at all surprised if they can put a nought or two on that number.

People keep saying this and it keeps not happening.

ChatGPT Pro was $200/mo when it launched in '23 for a ~100B class model with 8k context. Claude Max is now the same price for practically unlimited access to a ~1T class model with 1M context.

Moore's Law never died, it just switched architectures.

[swiftcoder]

Not to mention, the cost/performance of the baseline keeps falling. The cost-effective Deepseek V4 Flash is better than frontier models from a year ago, at a fraction of the cost.