LearnGrokThis is the expected behaviour of the technology rather than a fault in your account, and understanding why saves you from re-asking in five different phrasings.
Why it happens
Training has a cutoff. A model's knowledge ends at a date that precedes its release, which in turn precedes today. Anything that changed since is absent unless it was retrieved at request time.
It has no privileged access to its own product. A model cannot query the billing system, the feature flags or the pricing page. Asked about its own limits, it produces the most plausible answer from training: what was true once.
Retrieval is not omniscience. Even with live search, the answer is assembled from whatever pages were returned, and those pages may themselves be stale third-party summaries.
What to do instead
For anything about Grok itself (which models exist, what a plan includes, what something costs, whether a feature has shipped), go to the source:
- Models and API details: docs.x.ai
- Product and plans: x.ai
- Recent API changes: release notes
For everything else that is time-sensitive, ask for the source and open it. See Checking Grok's answers.
If it is confidently wrong about a fact you can check
Two useful moves:
- Ask in a new conversation. Long threads carry earlier mistakes forward, and a model that has committed to a wrong answer tends to defend it.
- Give it the current information in the prompt. Paste the relevant paragraph from the docs. Models are far better at reasoning over text you provide than at recalling text they were trained on.
What this page is not
It is not a claim that the answers are unreliable in general. The pattern is specific: questions about current state (dates, versions, prices, availability) are the weakest case, and they are also the ones people ask first.
What changes
Cutoff dates move with every model release, and retrieval behaviour differs between surfaces and plans. The advice does not: for current state, read the source.
