What Three Months of AI Would Have Cost on an API Key
4 min read
I run my development work through Claude Code on a flat subscription. So I don't actually get a per-token bill — but the meter is still running underneath, and I got curious what it would read if I paid à la carte. I added up every token from three months of transcripts and priced it at published API rates. The number is a good one to think with.
The headline: ~$19,000 a quarter
From April 8 to July 7, my AI tools moved 15.58 billion tokens. Priced at Anthropic's real per-model rates, that's $19,343 — roughly $6,400 a month.
Where the money goes, by model:
| Model | Metered cost | Share |
|---|---|---|
| Opus 4.8 | $12,819 | 66% |
| Fable 5 | $4,349 | 22% |
| Opus 4.7 | $830 | 4% |
| Opus 4.6 | $867 | 4% |
| Sonnet 4.6 | $452 | 2% |
| Haiku 4.5 | $26 | <1% |
Opus 4.8 does most of the work and takes most of the bill. Fable 5 is the expensive specialist — 22% of the cost on 13.5% of the tokens, because I reserve it for authoring the hard first drafts.
The real story is caching
Here's the number that matters more than the total: 90.5% of those 15.58 billion tokens were cache reads. Cached tokens bill at one-tenth of the input rate. That single fact is the difference between a viable workflow and an unaffordable one.
Run the counterfactual. Take the exact same work and turn caching off — every cached token billed at full freight:
- With caching: $19,343
- Without caching: $88,554
Prompt caching saved $69,000 — a 78% discount, 4.6× cheaper. The way an agent works, it re-reads the same large context on every turn: the file tree, the conversation so far, the system prompt. Cache that prefix and you pay for it once at 1.25×, then read it back all day at 0.1×. It's the boring infrastructure detail that makes agentic coding economically real. Without it, nobody's running 91,000 turns a quarter on consumer hardware.
What it'd cost somewhere else
Since the question was "what if I'd used an API key," here's the same 15.58-billion-token pool priced at each frontier lab's flagship. This is approximate — different tokenizers, different models, different cache economics — but the order of magnitude holds:
| Priced at… | Same workload |
|---|---|
| Anthropic Opus 4.8 | ~$17,600 |
| OpenAI GPT-5.5 | ~$16,500 |
| OpenAI GPT-5.4 | ~$8,300 |
| Google Gemini 3 Pro | ~$6,600 |
| Google Gemini 3.5 Flash | ~$5,000 |
You could run this workload three to four times cheaper on a mid-tier model. I don't, and the reason is the whole game: I'm not buying tokens, I'm buying correct output on the first pass. A model that's half the price but needs three tries to get a migration right isn't cheaper — it's more expensive and slower, and it burns my attention re-reviewing its work. The frontier tier earns its rate on exactly the tasks I hand it.
Two ledgers
There's a second bill this post doesn't show, and it's the interesting one: $0. Everything my homelab does overnight — the security reviews, the article scanning, the code review, the voice assistant — runs on local models on a Mac Mini, with no API anywhere in the loop. I wrote about why. That stack is genuinely free after hardware.
So I keep two ledgers. The metered one you just read is for the work where being wrong is expensive — interactive development, where I want the strongest model that exists. The $0 one is for the high-volume, fault-tolerant batch work where a local 9B model is plenty. The skill isn't picking the cheapest model. It's routing every token to the cheapest model that can actually do that job — which for a nightly RSS triage is free, and for authoring a migration is Fable 5, and the discipline is knowing the difference.