AI made us feel 20% faster but measured 19% slower. Here's the hidden variable nobody's tracking.
A study making rounds this week: developers using AI coding assistants felt 20% faster. Measured output? 19% slower.
The thread exploded. Hundreds of comments debating methodology, task complexity, cognitive offloading, attention fragmentation.
But I've been reading the whole thread and nobody — literally nobody — mentioned the variable I think matters most:
How much is your AI subscription costing you per hour of work?
The cognitive tax nobody measures
When you're paying $20/month for Claude or ChatGPT, that's not a rounding error in a developer salary. That's a decision.
For a developer in San Francisco making $150k/year, it's nothing. Use it all day. Don't think twice.
For a developer in Lagos making ₦800,000/month (roughly $500 USD equivalent in purchasing power), $20/month is 4% of take-home pay.
For a developer in Manila making ₱35,000/month, it's more than 3 days of income.
For a developer in Dhaka making BDT 40,000/month, it's 5.5% of gross salary.
When AI tooling costs that much, you don't use it freely. You ration it.
You save questions for "important" tasks. You don't iterate. You don't experiment. You don't use it to think through architecture problems — you use it to generate boilerplate you already know how to write, because that feels like you're "getting your money's worth."
Rationing = worse outcomes + higher cognitive load.
The study measured American developers with subsidized or employer-provided AI access. Nobody was rationing.
What the 19% slower number probably means
My guess — and I've spent time in developer communities in Nigeria, Indonesia, Philippines, and India — is that the productivity drag comes from trust miscalibration.
When you're paying premium prices, you second-guess more. You check every output. The tool feels like a $20/month employee you have to supervise.
When the price is negligible, you treat it like a rubber duck. You think out loud. You iterate fast. The tool becomes an extension of thinking rather than a line item.
The 19% slower finding might actually be a $20/month finding.
What affordable access changes
I've been running SimplyLouie — a $2/month Claude wrapper — for a few months now. The goal was specifically to test whether price point changes how developers use AI.
Some local pricing equivalents for context:
- India: Rs165/month (vs Rs1,600+ for ChatGPT direct)
- Nigeria: ₦3,200/month (vs ₦32,000+ direct)
- Philippines: ₱112/month (vs ₱1,120+ direct)
- Kenya: KSh260/month (vs KSh2,600+ direct)
- Indonesia: Rp32,000/month (vs Rp320,000+ direct)
- Brazil: R$10/month (vs R$100+ direct)
- Mexico: MX$35/month (vs MX$350+ direct)
- Bangladesh: BDT 220/month (vs BDT 2,200+ direct)
When the cost drops below psychological threshold — when it's genuinely less than your morning coffee — usage patterns change. Developers ask more half-formed questions. They experiment with prompts instead of trying to craft perfect ones.
That behavioral shift might actually close the 19% gap.
The question the study didn't ask
The 20% faster / 19% slower finding is real and worth discussing. But the study I want to see is:
Do developers with AI access below 0.5% of take-home pay show different productivity outcomes than developers with access above 2% of take-home pay?
My hypothesis: yes, significantly. Because the cognitive overhead of rationing expensive tools is itself a productivity tax that nobody's measuring.
The tool matters. The price matters more than we think.
Discuss
Do you change how you use AI tools based on what they cost? Do you notice yourself rationing prompts when you're on a paid plan vs. a free tier?
And for developers outside the US/EU: does the $20/month price point feel like a barrier to genuine, free experimentation — or have you found workarounds?
SimplyLouie is a $2/month Claude interface built specifically for developers who want full AI access without the subscription anxiety. 7-day free trial. Works everywhere.
Top comments (0)