Quick Answer: Somewhere a CFO is still saying "AI is too expensive to run." It was wrong last year and it’s adorable now. Stanford’s 2025 AI Index found the cost of GPT-3.5-level output fell about 280x in two years – from ~$20 to $0.07 per million tokens. Running the model is nearly free. The expensive part is the exception: the one wrong answer a human has to notice, chase, and fix. And here’s the joke – as tokens get cheaper, teams burn ~100x more of them, so the bill goes up while everyone celebrates the discount. The money lives in operations, not inference.
Cheap to generate, expensive to babysit
The line item that hurts isn’t the token – it’s the human time spent catching what the model got confidently wrong.
A 280x discount you spent 100x
Congratulations, you saved money into a bigger bill. Cheaper tokens don’t shrink spend; they unlock appetite.
Budget for the boring part
The cost you can’t see in the demo is the one that kills the rollout – which is the entire point of Managed AI: AI is an operating model, not a one-time build.
Proof it can go the other way
FAQ
Adapted from a LinkedIn post by Andrew Babkin, Managing Director & CTO at Azati – read the original. Stat: Stanford HAI, 2025 AI Index Report.
