Grok API rate limits: tiers, RPS caps and 429 errors explained

How Grok API rate limits work

SpaceXSI (formerly xAI) caps API traffic per model, using two numbers:

• RPS (requests per second): how many calls you can send each second. SpaceXSI derives it from a per-minute figure: RPS equals RPM divided by 48, with a minimum of 2.
• TPM (tokens per minute): how many tokens you can push through each minute. Cached tokens count toward TPM.

Limits apply to each model separately. A busy grok-4.7 workload does not use up your grok-4.3 allowance. When you go over a limit, the API returns HTTP 429 and the request is not processed.

Because cached tokens count toward TPM, caching lowers your bill but not your token budget per minute. Long prompts can hit the TPM cap well before you reach the request cap.

Spend tiers from $0 to $5,000

Your tier depends on how much you have spent on the API in total since January 1, 2026.

Four rules from the docs:

• Tiers rise automatically once you cross a threshold.
• Tiers never drop, even if your spending slows.
• Only prepaid credit purchases and paid invoices count.
• Higher limits for Grok Imagine image and video models go through SpaceXSI sales.

Example limits by model

The docs publish a per-model table. The figures below are what we read on October 18, 2026. The table was run together on the page, so treat these as approximate and check your console for the limits on your account.

Tiers 1 to 3 fall between these values. The console shows the exact limits that apply to your team.

The voice API has its own caps. The Grok Voice docs list 10 concurrent sessions per team and a 120-minute maximum session. See our Grok Voice API guide.

What to do when you hit a 429

A 429 means you sent more than your tier allows for that model in that window. It is not a billing error. A 429 is different from running out of credits, which causes requests to be rejected until you top up.

Standard practice for any rate-limited API applies here:

1. Retry after a short wait, and increase the wait after each failed attempt (exponential backoff).
2. Add a little random delay so parallel workers do not retry at the same moment.
3. Spread bursty jobs over time with a queue instead of firing them all at once.
4. Log which model and limit you hit. RPS and TPM problems need different fixes.

If you hit the TPM cap, shorter prompts help more than fewer requests. If you hit the RPS cap, combine small calls where you can.

Ways to get more from your limits

• Use the Batch API for work that can wait. Batch calls do not count toward rate limits, and they are 20% cheaper on grok-4.3 and the grok-4.20 models.
• Set a caching key. SpaceXSI strongly recommends prompt_cache_key or the x-grok-conv-id header. It will not raise your TPM, but it cuts the price of repeated input.
• Match reasoning effort to the task. Grok 4.7 defaults to high effort. Artificial Analysis found Grok 4.7 produced about 81k output tokens per task in its tests, more than double Grok 4.6, so heavy reasoning eats into token budgets fast.
• Pick the model with the right limits. In the published table, the grok-4.5 to 4.7 family starts with higher RPS than grok-4.3.
• Buy credits ahead of time. Since only prepaid purchases and paid invoices count toward tiers, a larger top-up moves you up sooner.

Enterprise and higher limits

Above Tier 4, SpaceXSI offers an Enterprise tier on request. Invoiced billing is also arranged through sales; it is off by default. Teams that need more Grok Imagine throughput must contact sales as well.

For business data, SpaceXSI says it does not train on business inputs or outputs and deletes them within 30 days unless it must keep them. Prices for every model are on our Grok API pricing page, and setup steps are in the Grok API overview.

Questions people ask