The window is the real one
Grok 4.5 reports a 500,000 token context window and that is the number you get. Unlike Codex there is no harness reservation to subtract, so a usage bar can use the reported figure directly.
Grok CLI guide: models and effort
Grok is the straightforward one on context. It reports the real window rather than an advertised figure you have to correct, and it compacts at a fixed proportion of it. The model list arrives in the opening handshake rather than from a separate call.
xAI
Grok 4.5 reports a 500,000 token context window and that is the number you get. Unlike Codex there is no harness reservation to subtract, so a usage bar can use the reported figure directly.
Auto compaction fires at 85 percent of the window. That is a point at which something happens, not a reduction in how much context you have, and treating it as a smaller denominator makes every usage readout wrong.
Grok reports its available models as part of the initial handshake rather than through a separate catalog call. A client already has the list before the first turn.
As with Codex, the levels belong to the model rather than to a fixed list, and they come from the same handshake.
An agent variant of a model, something like a build-specific name, generally shares the context window of the model it is based on. Its thinking support is a different matter and should not be assumed from the name.
Compare
| Claude Code | Codex CLI | Grok CLI | |
|---|---|---|---|
| Pick a model | Yes | Yes | Yes |
| Reasoning effort | Yes | Yes, per model | Yes, per model |
| Separate thinking budget | Yes | No | No |
| Where the level list comes from | The engine | The model catalog | The initialize handshake |
| Reports its context window | Yes | Yes, larger than usable | Yes, and it is accurate |
| Auto compaction | Yes | Yes | At 85 percent of the window |
Termdeck drives all three from one console, so the differences above stop being three separate habits to remember.
FAQ
Grok 4.5 reports 500,000 tokens, and that is the usable figure rather than an advertised one. There is no harness reservation to subtract, so the reported number can be used directly.
Auto compaction fires at 85 percent of the window. That is a threshold at which the conversation is compacted, not a smaller window, so the full 500,000 tokens remain the denominator for any usage readout.
The opening handshake. Grok reports its available models when a client connects rather than through a separate catalog request, so the list is in hand before the first turn runs.
Usually yes, because a variant generally inherits the window of the model it is based on. Whether it supports the same thinking controls is a separate question and should not be inferred from the name.
Claude Code, Codex and Grok in one browser console