Eventually, Programming in the AI Era Is About Tokens, Not Code
Claude just removed coding from the Pro plan.
Now if you want to actually build, you’re looking at the $100 tier.
Opus got smarter.
It also got ~30% more expensive in tokens.
This isn’t just pricing.
This is the end of the “AI trial phase.”
We’re officially in the cost phase.
After building with Claude recently, one thing became obvious:
Coding without AI is starting to feel… inefficient.
Not impossible.
Just wasteful.
And that’s where things get interesting.
I came across a case where a company spent ~$70k setting up an internal AI stack,
reduced their engineering team…
…and still ended up paying ~$20k/month in token costs.
They didn’t eliminate cost.
They just converted it.
From salary → tokens.
So the question is no longer:
“Should we use AI?”
It’s now:
“How much can we afford to burn?”
And this is where the role of developers starts to shift.
The most valuable engineers going forward might not be the ones who write the best code.
They might be the ones who:
- use fewer tokens
- say less
- get to results faster
Imagine this as an interview question:
“Build a login system within 1,000 tokens.”
That’s not a coding test.
That’s a cost-efficiency test.
You’d have to:
- design with minimal token usage in mind
- remove unnecessary context from prompts
- constrain the model to the shortest execution path
In other words:
You’re optimizing behavior under constraints.
Here’s the irony.
We’ve seen this before.
We already had a name for it:
Programming.
When it’s hot, you turn on the AC.
The solution exists.
But not everyone can run it all the time.
AI feels the same.
Right now, everyone is:
- vibe coding
- wiring up agents
- experimenting with MCP-like workflows
But this phase doesn’t last forever.
We’re moving from:
“Anyone can use it”
to
“Only those who can control it can afford to use it.”
And when that shift happens,
we’ll start seeing a real divide again.