Rate Limits

Rate limits, compared honestly

Almost no pricing comparison site publishes this, even though it's often what actually decides whether a production launch can handle traffic. Researched directly from each provider's own current docs — including the providers that don't publish real numbers at all.

As of Aug 25, 2026. This is hand-researched, not part of the automated daily pipeline — re-verify against the source link before relying on a specific number.

OpenAI

Spend-gated tiers

Gate

Spend-triggered, automatic (Free, Tier 1 at $5 paid, up to Tier 5 at $1,000 cumulative + 30 days)

Units

RPM, TPM, RPD, TPD — varies per model

Example

GPT-4o-class: ~500 RPM / 30K-200K TPM at Tier 1, up to ~10,000 RPM / 30M TPM at Tier 5

Self-check

Every API response includes x-ratelimit-remaining-requests/-tokens headers

Official docs ↗

Anthropic

Spend-gated tiers

Gate

Usage-history/account-standing based (Evaluation, Start, Build, Scale, Custom)

Units

RPM, ITPM, OTPM (input/output tracked separately), plus a monthly spend cap per tier

Example

Sonnet-class: Start tier ~1,000 RPM / 2M ITPM / 400K OTPM, up to Scale ~10,000 RPM / 10M ITPM / 2M OTPM

Self-check

Dedicated Rate Limits API to read your org/workspace's configured limits directly, plus anthropic-ratelimit-* response headers

Official docs ↗

Google

Spend-gated tiers

Gate

Keyed to total Google Cloud billing-account spend (Free, Tier 1 on billing enabled, Tier 2 at $100+3 days, Tier 3 at $1,000+30 days)

Units

RPM, TPM, RPD, plus a separate rolling 10-minute spend-based cap

Example

Free tier Gemini 2.5 Pro: 5 RPM / 100 RPD; Flash: 10 RPM / 250 RPD — Tier 1 jumps to roughly 150-300 RPM

Self-check

None published — AI Studio dashboard only

Official docs ↗

xAI

Spend-gated tiers

Gate

Cumulative API spend since Jan 1, 2026 (Tier 0-4 + Enterprise); tiers never downgrade once earned

Units

RPS (requests per second, not per minute) and TPM

Example

Fastest models: 150 RPS / 50M TPM at Tier 0 up to 500 RPS / 100M TPM at Tier 4; other models far lower (9-56 RPS)

Self-check

None published — xAI Console only

Official docs ↗

DeepSeek

Concurrency only, no RPM/TPM

Gate

Fixed per-model concurrency cap for every account, regardless of spend; higher limits require a manual capacity-expansion request

Units

Concurrent requests (no RPM/TPM ceiling at all)

Example

v4-pro: 500 concurrent; v4-flash/vision: 2,500 concurrent

Self-check

None published

Official docs ↗

Mistral

Spend-gated tiers

Gate

Auto-upgrading pay-as-you-go tiers at billing thresholds (>$20, >$100, >$500, >$2,000), plus a negotiated Priority/Enterprise tier

Units

RPS, TPM, and a tokens-per-month cap

Example

Tier thresholds are public; the exact per-tier RPS/TPM numbers are only shown in the logged-in dashboard, not a public docs page

Self-check

Account dashboard (admin.mistral.ai) only

Official docs ↗

Moonshot

Spend-gated tiers

Gate

Cumulative recharge/top-up amount, Tier 0 ($1) to Tier 5 ($3,000+)

Units

Concurrency, RPM, TPM, and TPD — whichever ceiling is hit first applies

Example

Tier 0: 1 concurrent / 3 RPM / 500K TPM / 1.5M TPD — Tier 5: 1,000 concurrent / 10,000 RPM / 5M TPM / unlimited TPD

Self-check

Console (platform.kimi.ai) only

Official docs ↗

Qwen

Fixed, same for everyone

Gate

Not spend-tiered — one fixed default limit per model per region, same for every account

Units

RPM and TPM per model

Example

e.g. Qwen3.8-Max in Singapore: 600 RPM / 1M TPM. Increases require a per-model, 30-day temporary support-ticket request

Self-check

None published

Official docs ↗

Z.AI

Not publicly documented

Gate

Tied qualitatively to Coding Plan tier (Lite/Pro/Max); described only as "dynamically adjusted based on resource availability"

Units

None published

Example

No RPM/TPM/concurrency figures published anywhere in official docs, even for the general (non-coding-plan) API

Self-check

None published

Official docs ↗

How to read this

Rate limits are not a level playing field across providers, and this page tries to show that honestly rather than force every provider into the same table shape. Five providers (OpenAI, Anthropic, Google, xAI, Moonshot) publish real, spend-gated tier ladders — start on a low tier and it rises automatically as you spend or build usage history. Qwen is different: its limits are fixed per model and identical for every account regardless of spend, with increases handled as one-off support requests. DeepSeek publishes no RPM or TPM ceiling at all — only a per-model concurrency cap. And Mistral and Z.AI both fall short of a fully public numeric table: Mistral's tier thresholds are documented but the exact per-tier numbers live behind a login, and Z.AI publishes no numeric limits anywhere in its official docs.

Self-checking your actual limit

Only Anthropic exposes a dedicated API to read your account's configured rate limits directly. OpenAI is the next-best option — every API response carries your current remaining-requests and remaining-tokens in its headers, so one cheap authenticated call tells you where you stand. Every other provider on this list requires checking a web console.

Why we don't auto-update this page

Unlike pricing, rate-limit tiers and thresholds aren't published in a form that can be safely scraped and re-published automatically — several providers gate the real numbers behind a login, and formats vary too much across the nine to normalize reliably. This page is refreshed by hand periodically; see the methodology page for which parts of the site are live-pipelined and which are dated snapshots like this one.

Frequently asked questions

Why doesn't this page have exact numbers for every provider?

Because not every provider publishes them. DeepSeek only publishes a concurrency cap, not RPM/TPM. Mistral's tier thresholds are public but exact numbers require a dashboard login. Z.AI publishes no numeric rate-limit information at all in its official docs. We report that honestly rather than guess or extrapolate.

How do I check my actual current rate limit?

For Anthropic, use their dedicated Rate Limits API. For OpenAI, make one cheap authenticated API call and read the x-ratelimit-remaining-requests and x-ratelimit-remaining-tokens response headers. For every other provider here, you'll need to check their web console — none of them publish a self-check API.

Why is this page dated instead of updated daily like the pricing table?

Rate-limit tiers change independently of pricing and aren't published in a consistently scrapable format the way per-token prices are — some providers gate exact numbers behind account logins. We'd rather show a periodically-verified, accurately-dated page than falsely imply this is live data.