Grok's 500K context window against the 1M-token models

Where 500,000 sits in the catalog

grok-4.6 and grok-4.5 both take 500,000 tokens of context and both emit up to 65,536 tokens. grok-composer-2.5-fast is smaller again at 200,000 in and 32,768 out. Those are the three Grok text ids, and 500,000 is the largest window the family offers here.

Several ids in the catalog take twice that. claude-sonnet-5, claude-opus-5 and claude-opus-4-8 all carry a 1,000,000-token window with a 128,000-token output ceiling. So does the whole Gemini line — gemini-3.8-flash, gemini-3.7-flash-high, gemini-3.6-flash-high, gemini-3-flash and gemini-3.1-pro-low each take 1,000,000 tokens, with a 65,536-token output ceiling (65,535 on the Pro id).

That makes the comparison concrete rather than abstract. Grok gives you half the input room of those ids, and on output it matches Gemini at 65,536 while the 1M Claude ids go to 128,000. If your prompts sit well under 500,000 tokens, none of this is a constraint you will ever hit.

What a large prompt actually costs on each

Window size is a limit; the bill comes from the per-token rate. Take a 150,000-token prompt, which fits everywhere mentioned here. On grok-4.6 input is billed at $0.20 per million after the 0.1x multiplier, so that prompt costs $0.03 — the same as gemini-3.1-pro-low at $0.20 per million, against $0.045 on claude-sonnet-5 ($0.30 per million), $0.075 on claude-opus-5 ($0.50 per million) and $0.0075 on gemini-3-flash ($0.050 per million).

Grok has a second tier, though, and it starts well before the window does. The listed Grok rates are the sub-200K prices: xAI doubles both input and output at or above 200,000 tokens in a single prompt. So the useful boundary on grok-4.6 and grok-4.5 is 200,000 tokens, not 500,000, and past it input is $0.40 per million and output $1.20.

That inverts the earlier comparison. A 400,000-token prompt costs $0.16 on grok-4.6 at the upper tier, against $0.12 on claude-sonnet-5 and $0.02 on gemini-3-flash, neither of which has a step in its rate card. Grok is the cheaper option below 200,000 tokens and the dearer one above it, which is a sharper distinction than the window size ever makes.

curl https://token-share.app/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN_SHARE_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "grok-4.6",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Summarise the attached module in five bullet points."}
    ]
  }' | jq '.usage'

Room is not recall

A 500,000-token window means a request of that size will be accepted. It does not mean the model weights every token equally, and it does not mean filling the window improves the answer. Padding a prompt is a straight cost increase with no guaranteed return — and on Grok it can be worse than linear, because crossing 200,000 tokens doubles the rate on the whole request.

The practical version of this: send what the task needs. If you are asking about three files, send three files. Pasting a whole repository to be safe multiplies the input line on every request in a loop, and an agent loop makes that a hundred requests, not one.

When a long prefix genuinely repeats across turns — a system prompt, a specification the conversation keeps returning to — cached reads on grok-4.6 are billed at $0.50 per million against the official rate, and at 0.1x that is $0.05 per million, a quarter of the $0.20 fresh-input rate. That is the lever for repeated context; a larger window is not.

Common questions