LLM API Rate Limits, Usage Tiers, and Quotas: A Practical Guide
Published 2026-08-02 · Updated 2026-09-27
Token price tells you what a request costs. Rate limits tell you whether your workload can run at the speed you need. The useful planning question is therefore two-part: estimate spend, then check whether your account's current request and token limits can sustain the peak load.
Separate rate limits, quota scope, usage tiers, and pricing
Providers enforce several limits at once. Common dimensions include requests per minute (RPM), tokens per minute (TPM), requests per day (RPD), and model- or modality-specific limits. A request can fail with a 429 response when the first relevant dimension is exhausted, even if the others still have headroom.
- Rate-limit dimension: the measured ceiling, such as RPM, input TPM, or output TPM.
- Quota scope: the account boundary that shares a limit, such as an organization, project, workspace, or model group.
- Usage tier: an account level that can change the limits available to you.
- Pricing or processing option: a separately priced mode such as batch or priority processing; do not assume it is the same thing as a usage tier.
Moving to a higher ordinary usage tier can raise capacity without changing the standard per-token price. Check pricing and capacity as separate inputs instead of treating a tier label as a price schedule.
OpenAI: organization and project limits with model-specific ceilings
OpenAI documents limits across dimensions including RPM, RPD, TPM, TPD, images per minute, and audio-specific limits. Limits are defined at the organization and project level, vary by model, and some model families share a common pool.
OpenAI usage tiers are tied to API spend and usually increase rate limits across most models as an organization graduates to a higher tier. For production planning, use the current limits shown in the account's limits view and the rate-limit headers returned by the API rather than a hardcoded table copied into application code.
Source checked 2026-09-27: OpenAI rate limits.
Anthropic: RPM, input TPM, and output TPM are separate
Anthropic documents Messages API limits in requests per minute (RPM), input tokens per minute (ITPM), and output tokens per minute (OTPM). Service limits apply at the organization level, while organizations can also configure lower workspace limits. Short bursts can hit a limit even when a one-minute average looks acceptable.
Input-token accounting is cache-aware for many Claude models, so cached and uncached input do not always consume ITPM in the same way. Anthropic exposes current configured limits in the Claude Console and, for organizations using the Admin API, through the Rate Limits API. That is safer than embedding community estimates that drift as tiers and model groups change.
Sources checked 2026-09-27: Anthropic rate limits and Rate Limits API.
Gemini API: limits are project-scoped, not API-key-scoped
The Gemini API commonly evaluates requests per minute (RPM), input tokens per minute (TPM), and requests per day (RPD). Google applies these limits per project rather than per API key, and the active values depend on the model and the project's usage tier.
Google directs developers to AI Studio for the project's active limits and notes that stated limits are not guaranteed capacity. Spend-based rate limits are also evaluated separately from RPM, TPM, and RPD. Vertex AI has its own quota system, so do not copy Gemini API limits into Vertex AI capacity planning.
Source checked 2026-09-27: Gemini API rate limits.
Capacity worksheet: convert peak traffic into RPM and TPM
Use the busiest minute you expect, not the monthly average. A simple first-pass capacity check is:
| Planning input | Formula | What to compare |
|---|---|---|
| Peak requests | Peak RPM = requests in your busiest minute | Provider/model RPM limit |
| Input tokens | Approx. input TPM = peak RPM × average input tokens/request | Input TPM or TPM limit |
| Output tokens | Approx. output TPM = peak RPM × average output tokens/request | Output TPM or TPM limit |
Example: 120 requests in the busiest minute with 2,000 input tokens and 400 output tokens per request implies about 240,000 input tokens/minute and 48,000 output tokens/minute. Compare every relevant dimension because the first limit you exceed becomes the binding constraint.
This worksheet is a screening tool, not a capacity guarantee. Burst enforcement, acceleration limits, concurrency, shared model pools, and modality-specific limits can still matter.
How rate limits affect cost planning
- Estimate spend: use the cost calculator with the workload's token volume and request pattern.
- Check peak capacity: convert the busiest minute into RPM and token-rate requirements, then compare them with the live account/project limits.
- Find the binding limit: plan around whichever dimension is exhausted first instead of relying on a single "TPM" number.
Do not assume that creating more API keys creates more capacity. OpenAI and Gemini document limits above the individual-key level, and Anthropic applies service limits at the organization level. Scale only through mechanisms the provider explicitly documents for the relevant account scope.
Strategies for working within rate limits
| Strategy | When it helps | Tradeoff or check |
|---|---|---|
| Queue + exponential backoff | Bursty traffic or temporary 429 responses | Adds latency; respect provider retry guidance and retry headers |
| Smooth traffic ramps | Rapidly increasing production load | Reduces burst/acceleration risk but requires admission control |
| Batch or asynchronous processing | Work that does not require an immediate response | Often has separate limits and pricing; not a real-time path |
| Prompt caching | Large repeated prefixes or context | Can change token accounting on some providers; verify model-specific rules |
| Request higher limits | Sustained legitimate demand near the current ceiling | Approval, account history, or spend requirements may apply |
| Model routing | A secondary model has suitable quality and independent documented capacity | Quality and pricing change; shared model pools can remove the expected headroom |
Pricing data sourced from our catalog. Check data sources for provenance and freshness.
Start here
Use the cost calculator to estimate your monthly volume, then check your target provider's rate limit documentation. If you're hitting limits, consider batch processing or model routing to stay within quotas.