LLM Price Table Drift

Summary

LLM price tables hardcoded into a codebase (per-model input/output/reasoning/cache-read rates) silently drift away from vendor-published pricing. The source documents two independent confirmations that yoyo's src/format/cost.rs misprices deepseek-v4-flash, proves the defect from the repo's own files plus external sources, and records prior art (yoagent's tests/price_audit.rs) showing that such drifts can survive 18 releases undetected. The design rule that emerges: a price-audit failure is a drift alarm that sends a human to the vendor page — never an auto-patch.

Key Points

• The defect is repo-provable. yoyo's .yoyo.toml comment (2026-09-15) states that deepseek-v4-flash is a legacy name served by DeepSeek-V4.1-Flash at the Flash price — i.e. it should be priced by the :200 arm. But cost.rs:187 matches v4-flash first (the table arms are substring tests) and prices it at (0.55, 0.0, 0.0, 2.19), a ~3.7× overstatement of the Flash input rate. The config comment and the pricing arm contradict each other without consulting any live page.
• Two external sources agree with the comment, not the arm. DeepSeek's own pricing page and models.dev/api.json both place deepseek-v4-flash at the Flash rate. So :187 is stale.
• The test pin is vacuous. cost.rs:1713 (a near-miss guard) asserts the wrong number byte-identically. A test that agrees with the table it guards cannot detect the table drifting from the vendor.
• Drift can be long-lived and expensive. yoagent shipped claude_sonnet_5 at Sonnet 4.6 rates (3/15 vs published 2/10) from v0.9.0 through v0.16.5 — 18 releases with a 50% overstatement on every cost_usd — found by a human asking, not by a mechanism.
• Cache-hit pricing is easy to get dangerously wrong. Cache-read rates are often an order of magnitude below cache-miss (DeepSeek: 0.003 vs 0.15 per 1M). Leaving at is not "conservative cheapness" — it is a wrong number that can report a bill than reality.