XAI Models Conetx Window Size Appears Or Is Actually Configured Smaller Than Available

[aosama]

Select Topic Area

General

Copilot Feature Area

VS Code

Body

When i added XAI models through my API key, i oberver that the context windows defined are way less than available from XAI. For example context window for Grok 4.1 Fast Thinking is 1 Million but in the IDE i see it as 120K ... is this by design or a gui bug or actually bug?

[ganapathijahnavi]

This is usually by design, not a bug.
Many IDE integrations (including VS Code extensions) cap the usable context window below the model’s theoretical maximum. Even if XAI advertises 1M tokens, the extension may:

1. Enforce its own safety/performance limit (e.g., 120K)
2. Limit context due to streaming/tooling constraints
3. Restrict based on current API tier or routing layer
4. Use a different variant than the full 1M-token version

The UI typically shows the effective context window available through the integration, not the model’s absolute maximum. If you call the model directly via XAI’s API (outside the IDE), you may see the larger context limit.
So this is most likely an integration limit, not a GUI bug.

[aosama]

But that doesnt give users of vscode a method to use a 1M context window model even through APIs --- looks like this is a hard handed restriction. I hope there is a pathway somewhere in Copilot VSCODE or COPILOT CLI to use those larger models 1M context through APIs.

[ganapathijahnavi]

Yeah, But If the model supports 1M tokens, it’s reasonable to expect you could use that through the API. The thing is, Copilot isn’t a direct pass-through to the provider. It adds its own routing, limits, and performance safeguards and those can cap the context window (like the 120K you’re seeing), even if the upstream model allows more. Right now there doesn’t seem to be a way in Copilot VS Code or CLI to override that limit. If you truly need the full 1M context, you’d likely have to call the XAI API directly outside of Copilot.
Hopefully GitHub exposes higher limits in the future, but at the moment it looks like an integration constraint rather than a bug.

[aosama]

Will open a feature request for the copilot team ----- something doesnt compute here -- other open source terminal based tools do not have that limitation to my knowledge --- but thanks a lot for the answers and your thoughts.

[aosama]

I added a feature request here microsoft/vscode#296046

[ganapathijahnavi]

Totally fair opening a feature request is the right call 👍
Some terminal tools are basically thin API wrappers, so they expose the full context. Copilot’s more of a managed layer inside VS Code, which is probably why it caps things. If enough people ask for larger-context support, there’s a better chance it gets prioritized.