Grok's Context Window Across Versions
8 min read · updated August 11, 2026
Every current Grok model has a six-figure context window and one of them has seven. The number that decides what you pay is not that one — it is 200,000, a threshold inside the window where the per-token rate doubles.
The documented windows
These are the context lengths published on xAI’s model list at the time of writing. Each model also appears on its own page with the same figure, its aliases and its region availability.
model context aliases grok-4.5 500,000 grok-4.5-latest, grok-build-latest grok-4.3 1,000,000 grok-4.3-latest, grok-latest grok-4.20-0309-reasoning 1,000,000 (dated snapshot) grok-4.20-0309-non-reasoning 1,000,000 (dated snapshot) grok-4.20-multi-agent-0309 1,000,000 (dated snapshot) grok-build-0.1 256,000 grok-build-latest
Two things in that table are worth more than the numbers. The first is that the newest model is not the largest window: xAI’s Grok 4.5 model page documents 500,000 tokens while the Grok 4.3 page documents 1,000,000. Context length is not a version number and does not increase monotonically across a family.
The second is that grok-latest currently resolves to grok-4.3 and grok-build-latest to grok-4.5 and grok-build-0.1 — that is, an alias can move you between models with a fourfold difference in window. If your prompts are anywhere near a limit, an alias is a liability rather than a convenience. That is the whole argument of pinning a dated snapshot.
The 200k step inside the window
xAI prices every text model in two tiers, and the boundary is a property of the request rather than of the model. The model pages state it plainly: requests whose prompt reaches 200k tokens are billed at the higher rate for all tokens in the request. Not the tokens above the threshold — all of them.
grok-4.3, per 1M tokens < 200k prompt >= 200k prompt input $1.25 $2.50 cached input $0.20 $0.40 output $2.50 $5.00 grok-4.5, per 1M tokens < 200k prompt >= 200k prompt input $2.00 $4.00 cached input $0.30 $0.60 output $6.00 $12.00
Work the edge case through, because it is the one that surprises people. A grok-4.3 request with a 199,000-token prompt and a 1,000-token answer is charged 199,000 input at $1.25/M and 1,000 output at $2.50/M — about $0.251. Add two thousand tokens of context and the same request becomes 201,000 input at $2.50/M and 1,000 output at $5.00/M — about $0.508. A one percent increase in prompt size doubles the bill.
So the operational context limit for a cost-sensitive workload is 199,999 tokens, not 1,000,000, and it is worth enforcing that in your own code rather than discovering it in a monthly invoice. The million is there for the requests that genuinely need it.
The threshold is measured on the prompt, which means it is measured on something you only partly control. Conversation history grows by itself, tool definitions are re-sent on every turn, and a single server-side search can add tens of thousands of tokens in one hop. A workload that sat comfortably at 150k in testing crosses 200k in production not because anybody edited a prompt but because a thread got long. The fix is a hard cap on history length in your own code, checked against the prompt_tokens the last response reported rather than against an estimate.
What is counted against it
The window holds the entire request, which on these models is more than text. xAI’s rate-limit documentation is the clearest statement of what consumes tokens: prompt tokens including text, image and audio; completion tokens; reasoning tokens on reasoning models; and cached prompt tokens, which still count even though they are billed at the reduced rate.
- Images. An image is charged as prompt tokens and reported separately in
usage.prompt_tokens_details.image_tokens. See how image input is counted. - Reasoning. On a reasoning model the internal trace occupies the window and is billed, whether or not it is shown to you. It is not free thinking space next to the window.
- Server-side tool results. Web and X search results are placed in the context before the answer is written, so a search turn can be far larger than the prompt you sent.
- Cached prefixes. A cached prefix is cheaper, not smaller. It counts fully against both the window and your tokens-per-minute limit.
The window is not the only ceiling
A request that fits the window can still be refused, because xAI limits throughput on two axes at once: requests per second and tokens per minute. Both are documented per model, and they do not scale together with context size.
model requests/sec tokens/min regions grok-4.5 150 50,000,000 us-east-1, us-west-2 grok-4.3 37 10,000,000 us-east-1, eu-west-1, us-west-2
Put the two numbers next to the context window and the tension is obvious. A single grok-4.3 request may hold a million tokens, and the documented per-minute allowance is ten million — so ten maximum-length requests exhaust a minute’s budget, well below the documented 37 requests per second. For long-context work the binding constraint is tokens per minute, not requests per second, and a concurrency limit tuned on request count will not protect you from it.
The categories that consume that allowance are the ones listed above, including cached prompt tokens — cheaper to buy, not smaller against the limit — and reasoning tokens. Exceeding either axis returns HTTP 429 Too Many Requests, and xAI’s guidance is exponential backoff. Tier is what moves these figures, and xAI documents tiers as determined by cumulative spend since 1 January 2026, from Tier 0 by default up to Tier 4 at $5,000 and enterprise on request.
Reading it at runtime
Do not hard-code the table above. The models endpoint is OpenAI-shaped and returns what the account can actually reach, which matters because availability is regional — grok-4.3 is documented in us-east-1, us-west-2 and eu-west-1, grok-4.5 in us-east-1 and us-west-2 only.
curl https://api.x.ai/v1/models \ -H "Authorization: Bearer $XAI_API_KEY"
And read usage back on every response rather than estimating. The response carries prompt_tokens, completion_tokens, total_tokens and a prompt_tokens_details object breaking the prompt into text, image, audio and cached components. Comparing that against your own estimate once, on real traffic, is worth more than any published figure — see why an offline GPT tokenizer will not give you that number.
Why this table goes stale
xAI retires model slugs on announced dates and redirects them. On 15 May 2026 eight slugs — including grok-3 and the whole grok-4-fast family — stopped resolving to their own models and began redirecting to grok-4.3. A redirect like that silently changes the context window under a running application, in that case from the retired model’s window to 1,000,000. That is documented on xAI’s retirement page, and it is the reason to treat any context figure — including this one — as a value to re-check rather than a constant to compile in.