Honest cost visibility

[jellologic]

Problem

Cost tracking for third-party models is broken across the whole ecosystem, and structurally so:

• Claude Code writes zero-token usage rows for non-Anthropic models (ccstatusline#147, maintainer: "This is a bug with non-Anthropic models in claude code"). Every JSONL-reading tool inherits this.
• ccusage's price lookup misses third-party model names entirely (#1201, #1473), and its fuzzy matching overcharges — #934 has gpt-5.4-mini matching gpt-5.
• Both requests to fix this were closed as not_planned (#656, #546).

Log-readers cannot solve this: the tokens aren't in the log.

Why we're positioned differently

We know the exact model id and its exact price per million tokens — we cached both at selection time. We don't have to infer either.

Proposal

Start honest and small: surface what a session would cost per million tokens at selection time, and estimate from whatever usage data is actually trustworthy.

Acceptance criteria

• Show input/output/cache pricing for the configured models
• Where usage data is available and non-zero, estimate spend
• Never present an estimate as measured — label it clearly
• Where the zero-token bug applies, say the data is unavailable rather than reporting $0.00
• No pricing shown for providers whose catalog lacks it

Explicitly out of scope for now

Full usage tracking. Until the upstream zero-token bug is fixed, any per-session total would be a guess. Reporting a confident wrong number is worse than reporting nothing — that's precisely the failure mode we're differentiating against.