MundoKit

LLM Rate Limit Calculator

Enter quotas or paste supported response headers to estimate sustained throughput and peak demand.

Inputs
Paste only x-ratelimit-* and Retry-After response headers. Recognized values override the example quotas below. Maximum 64 KiB and 1,000 lines.

Results

Planned sustained requests/minute
48
Reserved tokens per request
2,000
Mean in-flight requests at capacity
8
Peak demand status
Within planned capacity
Whole requests from remaining snapshot
No remaining headers provided
Effective quotas and demand
MetricValue
RPM quota60
TPM quota150,000
Peak token demand/minute60,000
  • Quotas are examples unless replaced by your account values or headers. Snapshot availability is not a burst guarantee. Short windows, shared quotas, daily limits and ramp controls can still reject requests. Mean concurrency is throughput × latency / 60, not a server concurrency limit. Reset and Retry-After values are displayed verbatim, not interpreted as a safe retry time.

Results update as you type. The share link stores your inputs after the # in the address, which is not sent to our server, but anyone you send it to can read them.

How it works

OpenAI documents token reservation as the greater of estimated request tokens and requested maximum output. A separate input-plus-output scenario is available for conservative planning.

capacity RPM = min(RPM quota, TPM quota ÷ reserved tokens) × (1 − headroom/100)
mean in-flight = capacity RPM × latency seconds ÷ 60

If a single reservation exceeds the effective token quota, executable capacity is zero: slowing submissions cannot make that request fit. Request quotas must be whole numbers. Headroom reduces planned throughput, not the per-request quota. Recognized response headers override corresponding quotas. A project token quota further caps TPM. Remaining snapshots give floor(min(remaining requests, remaining tokens ÷ reservation)). Missing snapshot dimensions are unconstrained in that snapshot estimate. These arithmetic scenarios assume minute quotas; check the actual header window. Values are unrounded until display (two decimals); snapshot request counts round down.

Worked examples

Default planning scenario

With 60 RPM, 150,000 TPM and 2,000 reserved tokens/request, 20% headroom gives 48 RPM. At 10-second latency, mean concurrency is 8. Peak demand of 30 RPM reserves 60,000 TPM.

Token-bound scenario

With 60 RPM, 30,000 TPM, 2,000 reserved tokens and 20% headroom, planned capacity is 12 RPM. Ten-second requests average 2 in flight.

Frequently asked questions

Should I paste authorization headers?
No. Paste only rate-limit response headers and Retry-After; no credentials are needed.
Does capacity guarantee no 429 errors?
No. Short windows, shared usage and rapid traffic increases can still cause rejections.
Is mean concurrency a user limit?
No. It estimates average requests in flight at the entered latency, not simultaneous users.
Do remaining headers describe sustained throughput?
No. They are a snapshot. Reset and Retry-After values are displayed without scheduling retries.

Sources

Last reviewed 2026-11-14; next review due by 2027-02-12. The method is written out above so you can check our math.