How cost is calculated

Six steps, each independently testable.

1. Normalise the model id

Lowercased, trimmed, and any bracketed context-window tag removed. claude-opus-5[1m] becomes claude-opus-5: the 1M window is the model’s standard context at its standard price, so the tag is a label rather than a rate.

2. Resolve it to a rate row

Longest-prefix matching, which is load-bearing because several entries are prefixes of others and priced differently:

Id	Rate	Prefix of
claude-opus-4	$15 / $75	claude-opus-4-8
claude-opus-4-8	$5 / $25	—
claude-fable-5	cache read $1.00	claude-fable-5-1
claude-fable-5-1	cache read $0.25	—

First-match or shortest-match would misprice all of them. The Fable pair is the nastier case: the two rows are identical in every other column, so a wrong match still produces a plausible-looking total.

A prefix only counts when what follows is a dated snapshot suffix — a dash and six or more digits. Without that rule an unreleased claude-opus-4-9 would quietly inherit retired claude-opus-4’s triple rate. Refusing to match makes it an unknown model, which warns loudly; guessing would not.

3. Pick the rate row for the speed

Fast mode is not a uniform doubling, and both directions of the mistake cost money. The table partitions models three ways:

• Supported — the fast row applies, with cache columns already stacked on the fast base rather than left at standard.
• Accepted but standard — the request runs at normal speed and bills at normal rates, so the fast row must not be used. Applying it would overcharge by 2x.
• Rejected — the API refuses fast for this model, so such a record should not exist. It bills at standard rates and raises a warning.

4. Multiply the five columns

cost = tokens ÷ 1,000,000 × rate

Applied independently to input, output, cache write 5m, cache write 1h, and cache read.

Why the cache tiers stay separate

A 5-minute cache write bills at 1.25x the base input rate; a 1-hour write at 2x. On the development corpus 96% of cache writes were 1-hour, so collapsing them into one figure would have understated cache-write cost by about a third.

cca prefers the nested cache_creation split in the transcript. When a record lacks it, everything is attributed to the 5-minute tier and the run is footnoted, because that is an approximation and you should know.

5. Apply modifiers

Modifier	Effect
inference_geo: "us"	1.1x on all five columns
Batch tier	0.5x on all five columns

An unrecognised region or tier is not guessed at. It bills at standard rates and warns — priority tier, for instance, costs more rather than less, so assuming parity would be wrong in the expensive direction.

6. Add server tools

Web search bills per request, so it sits outside the token multipliers. Web fetch is free beyond the tokens it pulls in.

Thinking tokens

Billed as output tokens and already inside output_tokens. cca never adds them.

A trap for anyone reimplementing this

Because thinking is summarised, the visible thinking text is far shorter than what was billed. A tool that counts output tokens from content blocks rather than from usage.output_tokens undercounts badly. cca reads the usage field.

Unknown models

Tokens counted, dollars excluded, model named in a footnote. cca never invents a rate.

How this is verified

Claude Code writes its own cost accounting into cost-state records. cca’s arithmetic reproduces one such record to within 1e-9, and three further assertions keep that from being a coincidence: the 5-minute cache attribution must not also match, adding thinking tokens must break the match, and claude-opus-5[1m] must resolve to the same rates as the bare id.