The same model ID is not the same price everywhere. A field guide.

Ask what a model costs and you will get a two-number answer: dollars per million input, dollars per million output. Building a pricing engine that has to match real invoices taught us that those two numbers are the output of a function with at least four arguments — and that every shortcut through the argument list produces a figure that looks right and is not.

The four components a pricing key needs#

Platform. The same model ships on multiple platforms at independently set rates, under independently shaped ids. Today's cross-platform Claude parity is a fact someone verified dimension by dimension, not a rule — and where a platform publishes no number at all, the honest state is "unpriceable there", not "same as elsewhere". AWS's own pricing page renders no table for several current Llama models; we keep that as a recorded checked negative, because "we looked and there is no number" and "nobody looked" must never collapse into the same absence.

The platform's own model id. Not a normalised family name. us.anthropic.claude-sonnet-4-5-20250929-v1:0 and claude-sonnet-4-5@20250929 are the same weights, but the id grammar carries platform-specific information — Bedrock's geo prefix encodes the endpoint class — and flattening ids is how customers' fine-tunes get priced as base models.

Endpoint class. Global versus regional versus multi-region, at +10% on the clouds for the regional forms — and on Bedrock the us./eu. inference-profile ids are what most production traffic actually uses, which makes "default to global" the single highest-frequency error available to a pricing engine. This component must be required with no default. It also has an asymmetric edge: a class can be unserved (Vertex publishes no regional rate for Opus 5 or Sonnet 5 because it does not serve them regionally), and an unserved class must refuse to price, not borrow the global figure.

Billing mode. On-demand is the baseline; batch is −50% where published (and only where published); premium "fast" tiers exist as real published prices on some models (a flat 2x on Opus — and only input/output, so the fast tier's cache economics are unpublished and a cached fast call must fail closed rather than get four invented numbers). Provisioned capacity — PTUs, GSUs, provisioned throughput — is not token-billed at all: per-token arithmetic there does not approximate the number, it fabricates one, and the only correct output is a refusal that says so.

The two components that resolve per request#

Two more dimensions cannot live in the key, because they are properties of the individual call.

Context tier. Several models re-price above a threshold — and re-price the entire request, not the overage. OpenAI's threshold is 272,000 input tokens, not the marketing page's rounded 270K; a rounded threshold re-prices requests the provider billed at standard. The threshold compares against the full prompt — uncached input plus cache reads plus writes — and where a long-context rate was never published for some dimension (cache writes, usually), a long-context call touching that dimension is unpriceable, honestly.

Time. Rates are validity windows, not constants. Promotions start and lapse on dates; Gemini's Flash line has a published 2x step on a calendar day; DeepSeek's rate varies by UTC wall clock, with peak weekday hours at double the off-peak price — so the same call costs different amounts at 09:00 and 11:00 UTC. A call must be priced at the rate in effect when it happened: resolving "now" against history silently re-prices the back catalogue every time a boundary passes. And a scheduled future rate is a claim about a world that has not happened yet — ours each carry a re-verify-by date that fails the build if it lapses before the boundary fires, because we once let a guessed price change fire unreviewed and charged three days of traffic 50% high.

The rules that keep the engine honest#

Four rules fall out, and they are the whole design. Exact key or nothing — an unknown combination prices to insufficient data, never $0, never a sibling's rate. No defaultable components — every shortcut above is someone defaulting an argument. Carry provenance — a rate read off a vendor page, one designated by contract, and one derived from a published rule ("cached input is 10% of input") are three different strengths of claim; an aggregate is labelled by its weakest member, and anything irreversible gates on the strong ones. Published-or-refuse for derived dimensions — where a discount is unpublished or zero (one fast host charges cached tokens at the full input rate), there is no saving to display, and rendering "$0.00 saved" would still be a claim about a discount that does not exist.

Every one of these rules exists because the shortcut it forbids produced, or nearly produced, a wrong number on a real dashboard. The rendered face of all of them is the dash, and the full keyed rate table — with each row's source and retrieval date — is what Vigil prices against.