Module cost 

AI cost attribution (DW-079): pricing tables and per-call cost computation.

The pricing table is a COMPILED view of the ai.pricing config map: provider model identifier -> per-1k-token micro-USD rates. It lives on the dataplane behind an ArcSwap and is swapped on every config reload (the same pattern as the AI budget engine), so a pricing change takes effect on the NEXT request without a restart.

Cost computation

cost_micros computes integer micro-USD for one call’s usage:

(input_tokens  - cached_tokens) * input_per_1k_micros  / 1000
+ cached_tokens * cached_input_per_1k_micros           / 1000
+ output_tokens * output_per_1k_micros                 / 1000

When batch is true, the batch rates (batch_input_per_1k_micros, batch_output_per_1k_micros) replace the standard input/output rates when declared; cached-token pricing is NOT combined with batch rates (a batch request uses the batch rate for all input tokens — the provider’s batch API does not report a separate cached tier). When optional fields are absent, the standard rates apply (the fallback).

Integer division truncates (the per-1k rate is the provider’s published unit; sub-micro-USD fractions are lost — the honest rounding boundary, pinned by tests). Saturating adds guard against overflow. An UNKNOWN model (no pricing entry) returns None — the caller applies the unknown_model_policy (allow / fail_closed / alert, DW-AI-03). The original fail-open behavior (cost 0) is preserved by the allow policy at the call site.

Dependency direction

ai depends on config only (see scripts/check_deps.py); this module reads config::ai::AiConfig and ai::types::Usage, both same-level or lower. The spend RECORD DTO lives in analytics (a plain struct, not importing ai::types::Usage) — the dataplane converts at the call site, keeping the dependency direction downward.

Structs

PricingTable

The compiled pricing table (DW-079): provider model -> rates. Built at dataplane refresh from the published config; immutable once built. Stored on the dataplane behind an ArcSwap and swapped on reload, so a pricing change applies to the next request with no restart.