LLM Rate Limit Calculator
On the Anthropic Start tier, Claude Sonnet 5.5 allows 1,000 requests, 2,000,000 input tokens and 400,000 output tokens per minute, so requests of 3,000 input and 500 output tokens run at about 667 per minute and 100,000 of them take 150 minutes; with 80% of input cached the rate rises to 800 per minute and 125 minutes (limits checked 2026-10-16).
How to use
- Pick a Claude tier preset or keep Custom and type the limits from your console. Enter 0 for a limit that does not apply.
- Enter the input and output tokens of one request, the share of input read from the prompt cache, and whether cached tokens are exempt from the input limit.
- Enter how many requests the job has and how long one request takes. Read the throughput, the binding limit, the job time, the concurrency and the pacing.
How it works
Requests per minute = the smallest of RPM, input TPM ÷ counted input tokens, output TPM ÷ output tokens and combined TPM ÷ (counted input + output tokens).
Counted input tokens = input tokens × (1 − cached share) when cached tokens are exempt, as on most Claude models; otherwise all input tokens count.
Job time = requests ÷ requests per minute. With a daily cap (RPD) below what the minute limits allow, each full day handles RPD requests and the rest runs on the last day.
Concurrency = requests per second × seconds per request, rounded up (Little’s law). Pacing = 60 ÷ requests per minute seconds between requests.
FAQ
Which limit usually binds first?
For long prompts it is usually the input token limit, for long answers the output token limit, and for short requests the request limit. The calculator names the one that sets the pace.
Do cached tokens count against Claude rate limits?
Anthropic documents that on most Claude models only uncached input tokens and cache writes count toward the input tokens per minute limit, so cache reads raise effective throughput. Claude Haiku 3.5 is an exception.
Why do I still get 429 errors below the limit?
Limits are enforced over short intervals with a token bucket, so a burst can exceed the limit even when the minute average is fine. Spread requests at the pacing shown and ramp up gradually.
Where are OpenAI and Gemini presets?
Their limits depend on model and account tier and are shown in each account dashboard, so the page leaves them as custom input instead of guessing. Copy the numbers from your limits page.