LearnGrokGrok moves fast enough that a page written today, including this one, is a snapshot. The durable skill is knowing where to look and what to re-check.
Where the truth lives
| Question | Where it is answered |
|---|---|
| Which models exist, and what they accept | docs.x.ai models |
| What changed recently in the API | docs.x.ai release notes |
| What a plan includes, and what it costs | x.ai pricing pages |
| Whether something is broken right now | xAI's status page, linked from x.ai |
| What the product does at all | x.ai/grok, x.ai/bot |
Notably absent: Grok itself. An assistant is a poor source on its own current version, limits and pricing, because that information post-dates its training and it has no privileged access to its own billing system.
What to re-check after a model update
- Prompts with a strict output format. Format adherence shifts between versions. If something downstream parses the output, re-run one real example.
- Anything tuned to a length limit. Input and output ceilings move.
- Cost assumptions. Per-token prices differ per model, and a default that silently rolls forward can change your bill.
- Refusal boundaries. Both directions: newly refused, and newly allowed.
Keeping an integration from rotting
Three habits, none of them exotic:
- Pin the model in code, not in your head. Read the model name from configuration so changing it is a one-line change, not a search across the repository.
- Keep one real request as a smoke test. Something with a known-good answer you can run after any dependency or model change.
- Log the model name with the response. When behaviour changes in production, the first question is always "which model produced this", and it is miserable to answer retroactively.
What changes
All of it, which is the point of the page. The table above is deliberately made of links rather than values.
