Traps
Named ways a headline price stops being the price. Each page is a definition, a computation, and the offender table — or an explicit we cannot say. We will not invent a ninth trap without a computation, and we will not print a tokenizer-tax percent.
- Context cliff computed· 60
Every published context-tier step function. Ladders keep every rung.
- Reasoning meter computed· 29
Separately billed reasoning tokens. The headline input/output pair is incomplete.
- Cache-write penalty we cannot say
The lean catalogue carries cached-input rates, not a cache-write tariff, so a break-even cannot be computed here.
- Credit cliff we cannot say
Included-pool exhaustion is a plan fact. This page will not guess a mid-month revert to API without a measured pool in the catalogue projection.
- Silent truncation we cannot say
No catalogue field records a hidden context cap (the ChatGPT Free Instant 27K claim is not in this payload). We will not guess.
- Alias drift computed· 18
*-latest is a moving SKU. The pointer is recorded; we do not freeze a vendor alias as a model.
- Precision bait computed· 84
Distinct sellers, not rows. A cheaper lower-quant endpoint is a trade, not a saving. See /spreads/.
- Tokenizer tax refused
The cross-vendor tokenizer-inflation figure is unverified in this repo. This page exists to refuse a percent until a public corpus exists. We will not print one.