A release note describes what the model can do. Four other documents decide what it costs you, and none of them appear in the announcement.
1. The price footnote
Gemini 4 Carbon launched at $2 and $10 per million tokens and moves to $4 and $20. The date is in a footnote on the pricing page, which is correct and easy to miss.
Read the pricing page on the day you budget, not the announcement.
2. The deprecation table
Model IDs retire. Every provider keeps a deprecation list with shutdown dates, and a product pinned to a retired ID fails at a date someone else chose.
OpenAI publishes a deprecations page. Anthropic and Google do the same. Put the shutdown date in your calendar the day you ship.
3. The retention exclusions
Zero data retention is arranged per organisation and covers a list of endpoints. Batch, files and code execution usually sit outside it.
If your data is someone else's, read the exclusion list before you read the price.
4. The benchmark setting
Anthropic's Opus 5.5 footnote is unusually specific: it names the effort level, the number of attempts, and the standard error of ±2.6 points. That is the rare case. Most tables give a score and a model name, and the score depends on settings the table does not show.
What candour looks like
The same release note says that at these capability levels, benchmark margins have become a less reliable guide to real-world differences. That sentence is worth more than the table above it, because it tells you what the numbers cannot.
Read the footnotes. They are where the bill and the caveats live.