| The next article initially appeared on Medium and is being republished right here with the creator’s permission. |
Don’t get me mistaken, I’m up all night time utilizing these instruments.
However I additionally sense we’re heading for an costly hangover. The opposite day, a colleague advised me a few new proposal to route 1,000,000 paperwork a day by way of a system that identifies and removes Social Safety numbers.
I joked that this was going to be a “million-dollar common expression.”
Run the maths on the “naïve” implementation with full GPT-5 and it’s eye-watering: 1,000,000 messages a day at ~50K characters every works out to round 12.5 billion tokens every day, or $15,000 a day at present pricing. That’s practically $6 million a yr to test for Social Safety numbers. Even in the event you migrate to GPT-5 Nano, you continue to spend about $230,000 a yr.
That’s a hit. You “saved” $5.77 million a yr…
How about working this code for 1,000,000 paperwork a day? How a lot would this price:
import re; s = re.sub(r”bd{3}[- ]?d{2}[- ]?d{4}b”, “[REDACTED]”, s)
A plain previous EC2 occasion might deal with this… A single EC2 occasion—one thing like an m1.small at 30 bucks a month—might churn by way of the identical workload with a regex and price you a couple of hundred {dollars} a yr.
Which implies that in observe, corporations will likely be calling folks like me in a yr saying, “We’re burning 1,000,000 {dollars} to do one thing that ought to price a fraction of that—are you able to repair it?”
From $15,000/day to $0.96/day—I do assume we’re about to see plenty of corporations understand {that a} pondering mannequin related to an MCP server is far more costly than simply paying somebody to jot down a bash script. Beginning now, you’ll be capable to make a profession out of un-LLM-ifying purposes.
