A crashing hook gets fixed the same morning, because something you expect stops arriving. This one arrived on time, every time, and said nothing had been spent. At 2,340 rows over 52 days the pipeline produced roughly 45 answers a day [14], and every one of them was wrong in the same direction.
The mechanism is worth being precise about, because it generalises past this one script. Version one was written on the assumption that the Stop hook's stdin carried token counts [6]. What stdin actually carries is session_id, transcript_path, cwd and hook_event_name, with no usage or model field anywhere in it [5]. Reading a field that is not there does not throw; the arithmetic still resolves, the append still succeeds, and the row still lands in ~/.claude/metrics/costs.jsonl [4]. A wrong assumption about a payload schema does not surface as an error. It surfaces as a plausible number.
Nothing in the setup was ever going to argue back. The author's stated reason for measuring at all is that Claude Code's MAX plan is a flat monthly fee with a soft ceiling, where heavy use chokes the following month's capacity [13]. That is exactly a situation with no itemised bill to reconcile against. The metric exists because no invoice enumerates sessions, which is also why a hollow metric survived seven and a half weeks. Self-reported telemetry with no counterparty is the case that most needs an assertion on the value, not just on the exit code, and zero is a legal cost, so absence and genuine idleness are indistinguishable without one.
Version two moves the read to the only thing the hook can trust: the transcript JSONL that Claude Code keeps writing during the session, parsed line by line for assistant turns carrying message.model and message.usage [7]. Input, output and cache token counts are summed across all turns, multiplied by per-model rates and divided by 1e6 [8]. That is a sound repair, and it inherits two soft spots.
The first is dispatch. getRates() looks for the substring 'haiku' or 'opus' in the model name and treats everything else as Sonnet [10]. With Sonnet input at $3.00 per million and Opus at $15.00 [9], a model name that matches neither string gets logged at one fifth of Opus input cost [16]. The default branch is a rate, not a failure, so this too would record quietly.
The second is the cache arithmetic the post uses to make its point. It argues that heavy caching on Opus, at $1.50 per million read tokens, can come out cheaper than Haiku's normal input [12]. The table printed a few lines above puts Haiku input at $0.80 per million [9], which makes the Opus cache read 1.875 times more expensive, not less [15]. The cache multipliers themselves are the real finding: writes at 1.25x input, reads at 0.1x [11], which is where the variance in an autonomous agent's bill actually lives.
