D1 — Oversized model

The problem

A frontier model answering short, mechanical requests — tag extraction, classification, routing glue — costs a multiple of what a one-tier-down model charges for the same job. Nobody notices, because each call is cheap; the route as a whole is not.

Detection

The detector flags model buckets whose median completion is short (default threshold: under 150 tokens) while running on a model that has a documented one-tier downgrade. The downgrade map is deliberate policy, not guesswork: exactly one tier down, never chained, never cross-provider. Frontier models without a ruled mapping produce an informational finding with no savings number instead of an invented one. Cached buckets are excluded — cached reasoning traffic is not evidence of oversizing.

Savings estimate

Bucket rows are re-priced at the suggested model's full four-rate card; the difference is the savings. This is linear token math — token means × rate delta — with no behavioral assumptions.

Worked example (golden fixture)

25 claude-opus-4-8 calls doing tag extraction: 1,500 uncached prompt tokens, 60 completion tokens each (below the 150-token short-completion threshold). Mapped downgrade: claude-sonnet-5 at its June 2026 introductory rates.

opus per call:    (1500 × $5  + 60 × $25) / 1e6 = $0.009
sonnet per call:  (1500 × $2  + 60 × $10) / 1e6 = $0.0036
observed savings: 25 × $0.0054            = $0.135
monthly (3-day window × 10):                $1.35

Confidence is estimated, and every D1 finding carries this caveat verbatim: "model suitability requires your own quality evaluation." We can prove the price difference; only you can prove the quality equivalence.

Known limitations

Savings assume the downgrade handles the workload — the report never claims that for you. Frontier models without a founder-ruled downgrade mapping are reported as informational only, with no dollar figure.