This tracker estimates OpenRouter-attributable infrastructure economics for every provider OpenRouter exposes.

Observed provider tokens come from each provider’s public OpenRouter page (“Tokens processed”) time series, including per-model breakdowns.

Endpoint prices come from OpenRouter model endpoint metadata and are snapshotted daily, including listed prompt, completion, and cache-read/write rates when published.

Cache / effective prices come from OpenRouter frontend stats (per-provider cache hit rate and observed effective input/output prices). These are stored as raw fields alongside listed prices — not forced into a single “truth” series.

Frontend billing toggle:

• Listed (default): listed prompt/completion × resolved input/output mix. Continuous across history; ignores cache discounts.
• Cache-adjusted: uses stored observed effective prices (or listed blended with cache-read × hit rate) when present; otherwise falls back to listed.

Both series are persisted. Humans can compare them and inspect raw listed in/out, effective in/out, cache hit rate, and token mix on provider model tables.

Token mix (prompt vs completion) for estimates follows:

1. provider × model observed input/output split, when available
2. else model-wide observed split from OpenRouter model activity (a proxy across all providers serving that model)
3. else assumed 80/20

Sensitivity cases (retained): prompt95 95/5, 90/10, 80/20, and 70/30.

Raw inferred estimates are retained alongside calibrated estimates. Methodology, pricing fields, token-mix source, and confidence are stored per provider/model/day.