I was spending $45/month on the Claude API. Not crazy money, but it bugged me because I knew most of my tokens were going to simple tasks that didn't need Opus-level reasoning.
here's what worked.
the problem
I was calling claude-opus-4-6 for everything. Refactors, typo fixes, code reviews, architecture decisions — all Opus, all the time. At $5/$25 per million tokens (input/output), those quick "fix this import" prompts were costing the same per-token as "design me a distributed cache invalidation strategy."
I looked at my usage and roughly 80% of my prompts were simple stuff. The kind of thing Haiku or Sonnet handles perfectly fine.
what I tried (and what failed)
Attempt 1: pin everything to Sonnet. Costs dropped immediately but quality tanked on the hard tasks. Multi-file refactors got confused. Architecture suggestions became generic. Sonnet is great but it's not Opus on the stuff that actually matters.
Attempt 2: manually switch models per task. This worked in theory but in practice I'd forget to switch back to Opus when I needed it. Or I'd second-guess myself: "is this task complex enough for Opus?" Decision fatigue killed it.
Attempt 3: route by task complexity. This is the one that stuck.
the routing approach
Simple rule: classify the task before sending it.
Quick edits, imports, typo fixes → Haiku at $0.25/$1.25 per M. These are 10-20x cheaper than Opus and the output is identical for simple operations.
Standard refactors, code reviews, test writing → Sonnet at $3/$15 per M. Handles 80% of real coding work at 40% the cost of Opus.
Architecture decisions, complex debugging, multi-system design → Opus at $5/$25 per M. Only when you genuinely need the reasoning depth.
my results after 30 days
Monthly spend went from $45 to $18. That's a 60% reduction. Quality on the hard tasks stayed the same because they still got Opus. I just stopped paying Opus prices for fixing semicolons.
the uncomfortable truth
Most of us are overpaying for AI because switching costs feel higher than they are. "What if Sonnet misses something?" is the fear. But after a month of routing, I can say: Sonnet doesn't miss anything on standard coding tasks. Haiku doesn't miss anything on simple edits.
The frontier tax is real. You're paying 10-20x more for capabilities you use 20% of the time.
what about opus 4.7?
The new tokenizer makes this even more relevant. Same prompts use 33-50% more tokens on 4.7 due to the tokenizer change. If you were on the fence about routing before, the 4.7 token inflation should push you over.
Route simple stuff to 4.6 (cheaper tokenizer), complex stuff to 4.7 (better reasoning). Best of both worlds.
try it
Start with the manual approach. Track your API calls for a week. Count how many are genuinely complex vs routine. I bet you'll find the same 80/20 split I did.
I'm a developer working on AI agent infrastructure. This is what I learned from actually looking at my token usage instead of just complaining about the bill.
Top comments (0)