I spent $788 on an AI coding agent in one day. Here's the breakdown.

I left an AI coding agent running for one day. Then I read the invoice.

$788. In about 13 hours.

I'm posting the real breakdown because I think a lot of people are quietly running up this kind of bill without seeing where it goes — and the fix is boring and effective.

The receipt

One day, 10:21–23:05. 11 sessions, 3,572 API calls across 4 models:

Model	Calls	Output tokens	Cache-read tokens	Cost
Fable 5 ($10/$50)	2,613	1.04M	448M	~$617
Opus 4.8 ($5/$25)	671	769K	248M	~$168
Haiku 4.5 ($1/$5)	242	27K	9M	~$1.70
Sonnet 4.6 ($3/$15)	46	6K	2M	~$0.90
Total	3,572			~$788

Two numbers reframed how I think about this:

• The flagship ate $617 by itself — 78% of the bill from one model I'd set as the default for everything.
• Haiku did 242 real calls for $1.70. A coffee. For work that, honestly, looked a lot like the work I was paying the flagship $0.24/call to do.

That's not a 2× or 3× gap. Per call it's a ~360× difference, and I was sending almost everything to the expensive end out of pure default-laziness.

What the cache-read column is telling you

Notice 448M + 248M = ~700M cache-read tokens. Agentic coding re-sends a big context every turn; cache reads are billed at ~0.1× input, which is the only reason this was $788 and not several thousand. The flip side: anything that breaks your cache (a changed timestamp, reordered tool list, a proxy that normalizes prompts) silently re-bills at full input price. On this volume, a broken cache is a 10× event.

The fix is routing, not abstinence

I didn't conclude "stop using good models." I concluded "stop sending everything to them." The pattern:

• Cheap model by default. Classification, file edits, boilerplate, retrieval — a fast small model handles these fine.
• Escalate on signal. Hard reasoning, ambiguous specs, failed attempts → bump to the flagship.
• Cap it. Per-key budgets so a runaway loop trips a limit instead of your card.
• Watch the cache. Keep the prompt prefix byte-stable so cache reads actually hit.

This is exactly what an AI gateway / model router does — it's the layer that lets you express "cheap by default, escalate when it's hard" once, instead of hard-coding a model everywhere. I've since taken the flagship out of the default path, and the same workload now lands in the low tens of dollars a day.

If you want the receipts

While digging into routing I built an open-source, pain-point-organized list of AI gateways — with a reproducible cost benchmark that prices concrete workloads (including a coding scenario with reasoning tokens) across 11 models, computed by a unit-tested script. Plug in your own token mix and see your real number before the invoice does:

github.com/cuihuan/awesome-ai-gateway · interactive cost tables

If you're running agents daily — have you actually looked at your per-model breakdown? I'd bet most of the bill is one model doing work a cheaper one could.