Six Months, Seven Models, and the Bill That Never Came
5 min read
I don't get a per-token bill. My development work runs through Claude Code on a flat subscription, so the meter runs somewhere I can't see it.
But I keep the receipts anyway. Every Claude Code session I run gets ingested into a Postgres table — 211,859 rows of turn-level telemetry, including the per-turn token counts the API reports back. So the bill is reconstructable. I just have to do the arithmetic.
I did the arithmetic for a quarter back in April. This is the whole ledger: January 19 through today, seven models, 38.18 billion tokens.
The number
Priced at Anthropic's published rates, six months of my development work comes to $33,159. Call it $5,400 a month.
That is the honest headline, and it is less interesting than the shape underneath it.
Token burn versus the money
Here is the thing me-in-January would not have guessed: the two bars don't line up.
Opus 4.8 is unsurprising — it does most of the work and takes most of the bill. Fable 5 is the interesting one. It's 6% of my tokens and 12% of my cost, because it's priced at double the Opus tier. I don't hand it routine work; I hand it the first draft of things that are hard to be wrong about. Two and a third billion tokens is a rounding error in volume and a real line item in dollars.
The blended rate across everything is $0.87 per million tokens. Opus 4.8's sticker input price is $5. That gap is the entire economic argument for agentic coding, and it has a boring name.
The boring name is caching
95.8% of those 38.18 billion tokens were cache reads. Not input. Not output. Re-reads of context the model had already seen — the file tree, the system prompt, the conversation so far — billed at one-tenth of the input rate.
Run the counterfactual. Same work, same models, caching off, every one of those tokens billed at full freight:
Caching saved $171,973 over six months. Without it this workflow costs 6.2× more, and I would not be running it.
There's a wrinkle I can't resolve from my own data: I'm pricing cache writes at the 5-minute rate (1.25× input). If some of those writes used the 1-hour TTL (2×), the real figure is closer to $38,600. My telemetry records the token counts, not which TTL wrote them. So read $33,159 as the floor and $38,622 as the ceiling.
Seven models, six months
The generational handoffs are visible in the monthly totals, and they are abrupt:
A few things I didn't expect to see:
July is 44% of the entire six months — 16.96 billion tokens in 22 days, and the month isn't over. That is not a pricing story, it's a workflow story. The agentic loop got good enough this summer that I stopped supervising it turn by turn.
Opus 4.7 lived for three weeks. 5.45 billion tokens between April 16 and May 5, then nothing. That's not a judgment on the model — 4.8 landed and I moved.
Some of the growth is a measurement artifact. Anthropic's pricing docs note that Opus 4.7 and later use a newer tokenizer that produces roughly 30% more tokens for the same text. Part of my curve going up is the ruler getting finer, not the work getting bigger. I only noticed this because the per-token cost went down while the token count went up.
The cheapest model was the most expensive
Per million tokens, actually spent:
| Model | Tokens | Cost | Effective $/MTok |
|---|---|---|---|
| Opus 4.8 | 22.93B | $19,606 | $0.86 |
| Opus 4.6 | 6.67B | $4,784 | $0.72 |
| Opus 4.7 | 5.45B | $4,362 | $0.80 |
| Fable 5 | 2.35B | $3,923 | $1.67 |
| Sonnet 4.5 | 528M | $258 | $0.49 |
| Opus 4.5 | 252M | $201 | $0.80 |
| Sonnet 4.6 | 10.4M | $25 | $2.40 |
Sonnet 4.6 is the cheapest model on that list by sticker price and the most expensive by effective rate — nearly 3× Opus 4.8. The reason is in the raw counts: 288 turns, and it wrote more cache than it read. 5.8 million tokens of cache creation against 4.5 million of cache reads. I used it so little that the cache never got a second look, so I paid the 1.25× write premium and never collected the 0.1× discount.
Caching isn't a discount on tokens. It's a discount on repetition, and a model you barely use has nothing to repeat.
What this doesn't count
Two ledgers are missing from that $33,159.
The first is output: 167.7 million output tokens, 0.44% of the total. Everything else was the model reading — its own context, my files, its previous turns. The thing that generates the code is a rounding error next to the thing that reads the code.
The second is the work that never touched an API. The overnight chain, the article scanning, the security reviews, the voice assistant — all of it runs on local models on hardware in my house, and the marginal cost is electricity. That ledger reads $0 and it isn't in any of these charts.
Which is really the point. The skill isn't picking the cheap model. It's knowing which of the two ledgers a given job belongs on — and then, on the metered one, making sure 96% of what you pay for is something the model has already read.