LLM API Rate Limits: One Reference for Every Provider
How rate limiting actually works on the five major LLM APIs — the documented tier structures, the per-model tables where a provider publishes them, what a 429 looks like on each, and the levers that raise your ceiling. Only documented facts: where a provider doesn't publish numbers, this page says so instead of guessing.
How each provider limits you — at a glance
| Provider | Limit dimensions | Tier structure | Per-model numbers published? | Official docs |
|---|---|---|---|---|
| Anthropic | RPM + input TPM + output TPM, per model. Cache reads don't count toward input limits. | Start → Build → Scale → Custom, advanced automatically. Monthly spend caps: $500 / $1,000 / $200,000. | Yes — full tier tables (see below) | Rate limits |
| OpenAI | RPM + TPM (+ RPD on some models), per model. | Tiers 1–5 by cumulative spend + account age ($5 → Tier 1 … $1,000 + 30 days → Tier 5). Manual increases open at Tier 5. | Partially — exact numbers are account-specific, shown on your platform Limits page. | Rate limits guide |
| Google (Gemini) | RPM + TPM + requests/day per model, plus a documented spend cap per rolling 10 minutes ($10 Tier 1, $200 Tier 2/3). | Free → Tier 1 (billing linked) → Tier 2 ($100 + 3 days) → Tier 3 ($1,000 + 30 days). | No — the docs point to your AI Studio dashboard for per-model numbers. | Rate limits |
| xAI (Grok) | Requests per second (derived from your RPM budget) + TPM, per model. | Tiers scale with cumulative team spend. | Partially — current per-model limits shown on the Console's Rate Limits page. | Consumption & limits |
| DeepSeek | Concurrency caps instead of RPM/TPM: 500 concurrent connections on V4 Pro, 2,500 on V4 Flash. | No spend tiers — capacity expansion by request. | Yes — documented caps | Rate limit |
Anthropic's published tier tables (the fully-documented case)
Anthropic is the only major provider publishing complete per-model, per-tier numbers. Limits are per model — traffic to one model doesn't consume another's budget — and cached input tokens don't count toward the input limit, so a high cache-hit rate multiplies effective throughput.
| Tier | Model | Requests / min | Input tokens / min | Output tokens / min |
|---|---|---|---|---|
| Start | Claude Fable 5 | 1,000 | 500,000 | 100,000 |
| Start | Sonnet 5 / Opus 4.x / Haiku 4.5 | 1,000 | 2,000,000 | 400,000 |
| Build | Claude Fable 5 | 2,000 | 1,500,000 | 300,000 |
| Build | Sonnet 5 / Opus 4.x / Haiku 4.5 | 5,000 | 5,000,000 | 1,000,000 |
| Scale | Claude Fable 5 | 4,000 | 4,000,000 | 800,000 |
| Scale | Sonnet 5 / Opus 4.x / Haiku 4.5 | 10,000 | 10,000,000 | 2,000,000 |
Source: Anthropic's official rate-limits documentation, retrieved July 14, 2026. Opus 4.x is a combined bucket across Opus 4.5–4.8; the Message Batches API has its own separate limits (1,000–4,000 RPM and a 200K–500K request queue by tier).
What hitting a limit looks like
- Every provider returns HTTP 429, usually with a
retry-afterheader. Anthropic exposes live budget state inanthropic-ratelimit-*headers (limit, remaining, reset per dimension); OpenAI does the same viax-ratelimit-*headers. - Limits refill continuously, not per calendar minute. Anthropic documents a token-bucket algorithm — a “1,000 RPM” budget can behave like ~16 requests per second, so short bursts can 429 even when your average is fine.
- Limits are enforced per model (per model class on some providers), so a fallback model is also a throughput strategy, not just a reliability one.
- Sudden traffic ramps can trigger separate acceleration limits — Anthropic documents 429s for sharp usage spikes even below your ceiling. Ramp up gradually.
Getting more throughput — the documented levers
- Exponential backoff with jitter on every 429, honoring
retry-after. Table stakes. - Prompt caching. On the Claude API, cache reads don't count toward input-token limits — Anthropic's own example: a 2M ITPM limit with an 80% cache-hit rate processes 10M total input tokens per minute.
- Batch APIs. OpenAI and Anthropic both run asynchronous batches under separate rate limits at 50% of standard price — move anything non-interactive there.
- Spend your way up. Tiers advance on cumulative spend (plus account age at OpenAI and Google) — if a launch is coming, put paid traffic through the account early rather than the week before.
Frequently asked questions
What is a 429 error and how should I handle it?
HTTP 429 means you exceeded a rate limit. Every major provider returns it, and most include a retry-after header telling you how long to wait. The standard fix is exponential backoff with jitter: wait, retry, and double the wait on each failure, adding a random offset so parallel workers don't retry in lockstep. Respect retry-after when present — retrying earlier just fails again.
Which LLM provider publishes its rate limits?
Anthropic is the most transparent: its docs publish full per-model tables (requests, input tokens, and output tokens per minute) for every tier. DeepSeek documents concurrency caps per model. OpenAI documents its tier structure but the per-model numbers are account-specific, shown on your Limits page. Google documents only spend-based limits — per-model numbers moved into the AI Studio dashboard. xAI shows per-model limits in its console.
How do I raise my rate limits?
Mostly by spending: OpenAI advances you through five tiers based on cumulative spend and account age, and manual increase requests open up at Tier 5. Anthropic moves organizations up automatically and takes increase requests from its Console Limits page. Google unlocks Tier 2 after $100 spent plus 3 days and Tier 3 after $1,000 plus 30 days. xAI tiers scale with cumulative spend. DeepSeek takes capacity-expansion requests for its concurrency caps.
How can I get more throughput without a limit increase?
Three documented levers. First, prompt caching: on the Claude API, cached input tokens don't count toward input-token-per-minute limits at all, so a high cache-hit rate multiplies effective throughput. Second, batch APIs: both OpenAI and Anthropic run batch requests under separate limits at half price. Third, spread traffic across models or providers — limits are almost always enforced per model, so a fallback model effectively doubles your ceiling.
Scope note: this reference covers documented rate limits only. Latency varies by region, load, and account and has no authoritative public source — so we don't publish latency numbers. Limits change; the linked official docs are always the source of truth.
Go deeper
Provider limits change. We track it daily.
Original AI briefings for founders, PMs, and builders — free, every morning.