A YC-backed startup recently shared findings from 75+ customer conversations across every industry. One theme dominated everything else:
"Everyone defaults to the most expensive model because nobody knows which to use."
This isn't a technology problem. It's a routing problem. And it's costing companies millions.
The Numbers Nobody Wants to Talk About
Here's what they found:
- CFOs have one line item for AI — not broken down by team, project, or agent
- Token costs get compared to cloud invoices in every finance meeting
- Uber burned its entire 2026 AI budget in four months and couldn't connect the spend to customer outcomes
- One fintech said per-use-case model selection was the single highest-impact request they had
The punchline? Every company they talked to wanted the same thing: a way to know which model fits which task. Nobody has it.
Why "Default to Max" Is Burning Money
When nobody knows which model fits which task, the rational choice is the most powerful one. Every time.
I've seen this pattern firsthand. Running 10+ AI apps, our Claude Code bills hit $10K/month before we traced the root cause:
60-70% of coding tasks don't need your most expensive model. They need the right model.
Here's the actual breakdown from our production workloads:
| Task Type | What Most Teams Use | What Actually Works | Cost Delta |
|---|---|---|---|
| Planning & Architecture | Fable 5 / Opus 4.8 | Opus 4.8 ✅ | 0% |
| Implementation | Fable 5 / Opus 4.8 | Sonnet 4.8 | -85% |
| Linting & Formatting | Fable 5 / Opus 4.8 | Haiku 4.8 | -95% |
| Code Review | Fable 5 / Opus 4.8 | Sonnet 4.8 | -85% |
| Test Generation | Fable 5 / Opus 4.8 | Haiku 4.8 | -95% |
The waste is structural, not accidental.
"Just Pick a Cheaper Model" Doesn't Work Either
I hear this a lot. "Just use Sonnet for everything."
It breaks on architecture tasks — you get shallow scaffolding instead of thoughtful system design. "Just use Haiku for everything." It misses critical nuance in code review.
The real answer is task-level routing: automatically matching each coding task to the right model based on what's actually being done.
We built this into our workflow. The result: $10K/month → $3K/month. Same output quality on the metrics that matter — test pass rates, PR review accuracy, architecture coherence scores.
No model downgrade. Just smarter dispatch.
The Maturity Gap Is Enormous
The same research highlighted another pattern: the gap between companies who've figured this out and those who haven't is widening fast.
- One neobank moved AI into mission-critical ops and is running a hiring freeze against 50-60% growth — funded entirely by AI efficiency gains
- Meanwhile, a Director of AI at a large agency said they're "3-4 years from ready"
The difference isn't budget or talent. It's visibility. The neobank knows exactly which models power which workflows, what they cost per task, and where the waste sits. The agency treats AI as one amorphous blob of compute.
What Actually Works (From Cutting Our Own Bill by 70%)
Four things, in order:
1. Audit your model usage by task type.
Most teams have genuinely never done this. You'll be shocked how much Opus or Fable is being used for tasks that Haiku handles perfectly.
2. Identify your "Opus tasks" vs "Haiku tasks."
In our portfolio, it's roughly 30/70. Yours might differ — the point is that the split exists.
3. Route automatically.
Manual model selection doesn't scale past a single developer. You need a system that classifies the task and dispatches to the right model without human intervention.
4. Measure per-task quality.
Cost savings mean nothing if output degrades. Track the metrics that matter for each task type — not just "did it complete" but "did it complete well."
The Bottom Line
The companies winning the AI cost game aren't the ones with the best prompts, the biggest budgets, or the newest models.
They're the ones who figured out that model selection is a routing problem, not a spending problem.
The CFO in that research wanted AI spend tracked as precisely as ad spend — every dollar attributed, every channel measured. That same discipline is coming to AI. The question is whether you build it proactively or get surprised by the bill.
I run a portfolio of 10+ AI-powered apps and write about what actually moves the needle on AI costs. Previously cut our team's AI coding bills from $10K to $3K/month with task-level routing. Find me on X @aplomb2.
Top comments (1)
Routing is where AI cost becomes an operating problem. The expensive mistake is sending every step to the strongest model instead of deciding which parts need reasoning, which need retrieval, and which need deterministic code.