26/05/2026
Over the last few weeks, everyone has been talking about Uber “tokenmaxxing” its way through the 2026 AI coding budget in four months.
At GNH India, we’ve taken the opposite path.
Instead of chasing *more* tokens, we focused on **more logic per token**:
- We routed work through OpenRouter so we could choose the *right* model for each task, not the most expensive one.
- We separated deterministic workflows (rules, lookups, checks) from true reasoning tasks, so we don’t waste premium reasoning tokens where simple logic will do.
- We defined success as **time saved, errors reduced, and throughput improved per rupee**, not “tokens consumed” or “AI usage hours.”
- We built internal guidelines: if the task is repeatable and stable, automate it; if it’s rare and judgment-heavy, then spend tokens—deliberately.
The result: we’ve kept our AI costs under control *without* slowing innovation, because the real problem was never “too many tokens.”
The problem is **illogical usage of tokens**.
Tokenmaxxing is a phase.
Disciplined orchestration is a strategy.
If you’re a founder or CXO experimenting with AI, I’d love to compare notes on:
- How you measure “value per token” in your org
- How you decide which tasks truly deserve expensive models vs lightweight ones.
A top Uber exec said AI is not giving the company bang for its buck
