What is training data? The cutoff in the weights

By

Training data is what a model learned from, and it stops at a cutoff. Dates on the Grok 4.7, GPT-6.1 Terra and Claude cards, and how an agent reads your repo.

Training data is the text and other material a model learned from. It has a cutoff, so the weights include nothing newer than that date.

A coding agent can still read your repo, or search, during the session. That text is in the request. It is not in the weights.

Build an LLM from scratch is the training loop. This entry is the cutoff, and the dates the model cards actually print.

How training data works

Training reads a corpus and sets the weights. After that, the weights do not update when a library ships a new release or a doc page changes. The context window can still hold a file the agent just read. That is this session, not a new training run.

Training data and a later questionTraining fixes the weights up to a cutoff. A later question is answered from those weights unless a tool brings in a file or a search result.Training textWeights, up to thecutoffAnswerFile or search

Fine-tuning is a later, smaller training run on your examples. It does not move the original cutoff forward to today.

Training data in each tool

The cutoff belongs to the model, not to the coding app. Two of the four product pages do not print one.

Claude Code Codex Cursor Grok
A date on the product page Not on the model config page Not on the Codex models page Not on the models and pricing page Grok 4.7: May 2026
Where the date is The Claude models overview, for the current API ids The API card for the model you selected The card for the model you selected The models index

Claude Code

The model config page does not print a knowledge cutoff. The Claude models overview does, for the current API ids. In November 2026 both the reliable knowledge cutoff and the training data cutoff are Jun 2026 for claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5, and claude-haiku-5-5.

Those two rows match on that table. Don’t copy the dates onto an older model id the table does not list in those columns. See the models overview.

OpenAI Codex

The Codex models page does not print a cutoff. The API card does, per model. The gpt-6.1-terra card says Jun 01, 2026. GPT-6.1 Terra says the same date. See the Terra model card.

Cursor

The models and pricing page does not print a cutoff. Cursor hosts other companies’ models, so the date is the one on that model’s card.

The agent page does document a way to see something newer. The Web tool generates search queries. Read opens files in the repo. See the agent overview.

Grok

The models index says Grok has no knowledge of current events or data beyond what was present in its training data. The note under that section says the knowledge cutoff of Grok 4.7 is May 2026.

For anything newer, the same page says to enable server-side search tools, Web Search or X Search. See the models index.

Details checked against each vendor’s documentation in November 2026.

A real example

Ask for the file, not for a memory of the file:

Don't answer from memory. Read package.json and tell me which version of astro this project requires.

The weights may have seen some package.json during training. They have not seen this repo’s file unless the agent reads it. The same move works for a changelog, a migration guide, or a model card that was published after the cutoff.

When to use the cutoff

Use it when the question is “does this model already know about a release?” If the release is newer than the date on the card, point the agent at the doc or the file.

Use a search tool when the fact is a current event. Grok’s models page is explicit: without Web Search or X Search, the model does not have that data.

The course lesson What a language model does is the beginner version of where training text comes from.

Pitfalls

One date does not cover every model. Grok 4.7’s card says May 2026. Terra’s card says Jun 01, 2026. The four Claude ids above say Jun 2026. A model whose card you have not opened does not inherit those dates.

A search result is not the model remembering the news. It is text the tool put in the request. Drop the tool and the cutoff applies again.

The cutoff is not the date the coding tool’s docs were written. Claude Code can be newer than the weights of the model it is calling.

  • Fine-tuning: Fine-tuning trains an existing model a little more on your examples, for a narrow job. OpenAI's platform is closed to new users. The local path is posted here.
  • Context window: The context window is everything a coding agent can see at once. How to check it, free it and keep it small in Claude Code, Codex, Cursor and Grok Build.
  • Multimodal model: A multimodal model takes more than text, such as a screenshot, in the same request as the prompt. How Claude Code, Codex, Cursor and Grok accept an image.

All AI glossary entries →

Want me to talk about your product? You can sponsor this site.