Every LLM API provider throttles traffic, but each one measures limits in different units, applies them at a different scope and moves customers between tiers on different rules, so a workload that runs cleanly on one API can fail with HTTP 429 on another. This page compares the documented rate limit models of eight API providers: what they measure, whether limits sit on the organisation, project, team or key, how usage tiers are earned, what the error and headers look like, and how batch traffic is treated. It is written for developers sizing production workloads, platform teams running several providers behind a gateway, and buyers who need the published numbers in one place. Providers are listed alphabetically and order is never a ranking. Numbers are copied from each provider's own documentation; where a provider publishes nothing for a field, the page says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Limits change often and many providers show per-model values only inside their console, so confirm current values there before you plan capacity.
Key takeaways
- The units differ. Anthropic splits input and output tokens per minute, Google, Meta and OpenAI count requests and tokens per minute, xAI counts requests per second and tokens per minute, DeepSeek limits concurrent connections, and Cohere publishes requests per minute per endpoint.
- The scope differs. Google applies limits per project, Meta and xAI per team, Mistral AI per workspace, DeepSeek per account, and OpenAI at both organisation and project level; none of the pages checked applies the main limit per API key.
- Tiers are earned by spend, history or a key type. OpenAI, xAI and Google move you up on cumulative payments (Google adds waiting periods), Anthropic places organisations automatically on usage history and account standing, and Cohere separates free evaluation keys from paid production keys.
- Caching changes the maths on some APIs only. Anthropic's default input limit counts only uncached input tokens, while xAI states that cached prompt tokens count toward tokens per minute.
- Spend caps can look like rate limits. Anthropic, Google and OpenAI each document a monthly or windowed spend ceiling per tier that can return a 429 even when per-minute limits have headroom.
What is an LLM API rate limit?
A rate limit is the maximum volume of requests or tokens a provider accepts from one customer scope within a time window. Providers use them to share capacity fairly and to protect against runaway or abusive traffic. A usage tier is the bundle of limits a customer is currently assigned; moving up a tier raises the limits. Rate limits are separate from spend caps (a ceiling on money spent per month or per window) and from credit balance (prepaid money remaining), although all three can stop requests. For the status codes themselves and how to tell them apart, see the LLM API 401, 403 and 429 Errors Guide. This page focuses on the published limit models.
Rate limit models at a glance
As of 2026-11-15, from each provider's rate limit documentation.
| Provider | What is measured | Scope | Tier names | Where per-model values are shown | Source |
|---|---|---|---|---|---|
| Anthropic | Requests per minute (RPM), input tokens per minute (ITPM), output tokens per minute (OTPM) | Organisation | Evaluation, Start, Build, Scale, Custom | Docs tables and the Rate limits page in Claude Console | Rate limits |
| Cohere | Requests per minute per endpoint (inputs per minute for Embed); 1,000 calls a month on evaluation keys | Key type (evaluation or production) | Evaluation (trial) key, production key | Docs tables | Rate Limits |
| DeepSeek | Concurrent connections per model | Account, not individual keys | Standard and expanded-quota accounts | Docs page | Rate Limit |
| Google (Gemini API) | RPM, input tokens per minute (TPM), requests per day (RPD); IPM or TPD on some models | Project, not key | Free, Tier 1, Tier 2, Tier 3 | AI Studio rate limit page | Rate limits |
| Meta (Model API) | RPM and TPM per model; stream limits for transcription | Team, not key | No tier names stated | Docs table | Pricing and rate limits |
| Mistral AI | Requests per second (RPS), tokens per minute and per month | Workspace | Free API tier and higher usage tiers (names not stated on the page) | console.mistral.ai/limits | Rate limit and usage tiers |
| OpenAI | RPM, RPD, TPM, TPD, images per minute, audio minutes for some streaming models | Organisation and project | Free, Build, Launch, Grow | Not listed on the guide page checked | Rate limits |
| xAI | Requests per second (RPS), TPM | Team, per model | Tier 0 to Tier 4, Enterprise | Models page in xAI Console | Rate limits |
How do usage tiers work at each provider?
As of 2026-11-15. Thresholds are cumulative unless stated; "cap" is the provider's own spend ceiling for the tier.
| Provider | How you move up | Tier thresholds and caps | Source |
|---|---|---|---|
| Anthropic | Automatic, "based on usage history and account standing"; new or low-history organisations may start in Evaluation | Start: $500 monthly spend cap. Build: $1,000. Scale: $200,000. Custom: no monthly cap, limits arranged with the account team | Rate limits |
| Cohere | Create a production key on the API keys page; contact support for increases | Evaluation keys free but limited to 1,000 API calls a month; production keys paid | Rate Limits |
| DeepSeek | Capacity expansion form; DeepSeek states there is no additional cost for expansion | No spend tiers stated | Rate Limit |
| Google (Gemini API) | Cumulative Google Cloud spend on the linked billing account plus days since first payment; Free to Tier 1 is instant, later upgrades within 10 minutes | Free: active project or free trial. Tier 1: active billing account linked, cap $250. Tier 2: $100 paid and 3 days, cap $2,000. Tier 3: $1,000 paid and 30 days, cap $20,000 to $100,000+ | Rate limits |
| Meta (Model API) | Not stated by Meta on the page checked | Single published limit set per model | Pricing and rate limits |
| Mistral AI | Contact support with your use case; a higher tier is recommended for production | Not stated by Mistral AI on the page checked | Rate limit and usage tiers |
| OpenAI | Automatic as total credit purchases reach each threshold | Free: allowed geography, $100 a month. Build: $5 purchased, $500 a month. Launch: $100, $5,000 a month. Grow: $500, $200,000 a month | Rate limits |
| xAI | Automatic on cumulative API spend since 1 January 2026 (prepaid purchases or fulfilled invoices); tiers never downgrade | Tier 0: $0. Tier 1: $50. Tier 2: $250. Tier 3: $1,000. Tier 4: $5,000. Enterprise on request | Rate limits |
Google also documents a separate spend limit on a rolling 10-minute window: $10 at Tier 1, $50 at Tier 2 and $200 at Tier 3, returning 429 RESOURCE_EXHAUSTED when exceeded.
Published limits for current models
As of 2026-11-15. Only values each provider prints on its public page and clearly labels are shown. Several providers publish per-model numbers only in their console; those rows say where to look.
| Provider | Model or endpoint | Published limit | Notes | Source |
|---|---|---|---|---|
| Anthropic | Claude Opus 5.5 and Claude Sonnet 5.5, Start tier | 4,000 RPM, 2,000,000 ITPM, 400,000 OTPM | Opus 4.x models share one combined limit | Rate limits |
| Cohere | Chat with Command A | Trial 20 requests a minute; production 500 a minute | Command A+, Reasoning, Translate and Vision production limits via sales | Rate Limits |
| DeepSeek | deepseek-flash; deepseek-v4-pro | 2,500 and 500 concurrent connections | A request counts from sending until the response completes | Rate Limit |
| Google (Gemini API) | Per model | Shown in AI Studio, not on the docs page | Batch enqueued tokens for Gemini 2.5 Pro: 5,000,000 (Tier 1), 500,000,000 (Tier 2), 1,000,000,000 (Tier 3) | Rate limits |
| Meta (Model API) | muse-spark-1.3, muse-wonder-1.0, muse-ember-1.0 | 3,000 RPM, 4,000,000 TPM each | Contributor variants: 150 RPM, 3,000,000 TPM | Pricing and rate limits |
| Mistral AI | Per model | Shown at console.mistral.ai/limits | Free tier described as having restrictive limits for testing | Rate limit and usage tiers |
| OpenAI | Per model | Not listed on the guide page checked | Rate limit guide gives tier rules, not per-model values | Rate limits |
| xAI | grok-4.7 (also grok-4.5, grok-4.6) | TPM 50M (Tier 0), 53M, 60M, 74M, 100M (Tier 4) | RPS derived from RPM / 48, minimum 2 | Rate limits |
How each provider counts tokens and bursts
- Anthropic: uses the token bucket algorithm, so capacity refills continuously rather than resetting on the minute, and a 60 RPM limit may be enforced as 1 request per second. Only
input_tokensandcache_creation_input_tokenscount toward ITPM;cache_read_input_tokensdo not. The documentation's example: with 2,000,000 ITPM and an 80 percent cache hit rate, about 10,000,000 total input tokens a minute can be processed. OTPM counts tokens actually generated;max_tokensis not counted. A sharp traffic increase can triggerramp_limit_exceededeven below the limits; the advice is to increase by at most one step per minute. - Cohere: limits are per endpoint and per model family, for example Rerank at 10 requests a minute on trial keys and 1,000 on production keys, and Embed at 2,000 inputs a minute on both.
- DeepSeek: counts open connections rather than tokens. While a request waits, the server sends empty lines (non-streaming) or
: keep-alivecomments (streaming), and closes the connection if inference has not started after 10 minutes. On expanded-quota accounts eachuser_idgets its own limit. - Google: exceeding any one of RPM, TPM or RPD returns an error; RPD resets at midnight Pacific time.
- Meta: requests succeed again as soon as usage falls below the limit. Background submissions (
background: true) are capped at 600 a minute per team; running background responses continue. - Mistral AI: combines a per-second request limit with per-minute and per-month token limits per workspace.
- OpenAI: unsuccessful requests still count toward per-minute limits. Traffic reaching 1 million input TPM should grow by no more than 50 percent every 15 minutes.
- xAI: bursts are capped at roughly one forty-eighth of the minute's request budget. TPM counts prompt, completion, reasoning and cached prompt tokens; cached tokens are billed at a lower rate but still count.
What happens when you hit a limit?
As of 2026-11-15.
| Provider | Error returned | Retry guidance in docs | Rate limit headers | Source |
|---|---|---|---|---|
| Anthropic | 429 naming the limit; enforced_spend_limit_reached for tier spend cap (no retry-after); user-set spend limit returns 400 | Wait the retry-after seconds | anthropic-ratelimit-requests-*, -tokens-*, -input-tokens-*, -output-tokens-*, plus anthropic-priority-* on Priority Tier | Rate limits |
| Cohere | Not stated on the page checked | Contact support for an increase | Not stated by Cohere on the page checked | Rate Limits |
| DeepSeek | 429 when the concurrency limit is exceeded | Links to a separate error codes page | Not stated by DeepSeek on the page checked | Rate Limit |
| Google (Gemini API) | 429 RESOURCE_EXHAUSTED | Request an increase through the paid tier form, with no guarantee | Not stated by Google on the page checked | Rate limits |
| Meta (Model API) | 429; transcription streams close with code 1013 | Exponential backoff with jitter from about 500 ms; honour Retry-After | x-ratelimit-limit-tokens, -remaining-tokens, -limit-requests, -remaining-requests | Pricing and rate limits |
| Mistral AI | Not stated on the page checked | Contact support to raise limits | Not stated by Mistral AI on the page checked | Rate limit and usage tiers |
| OpenAI | 429; 503 server_is_overloaded for model overload | Treat Retry-After as a minimum and add jitter; otherwise exponential backoff; do not retry quota or billing errors | x-ratelimit-limit-*, -remaining-*, -reset-* for requests, tokens and project tokens | Rate limits |
| xAI | 429 Too Many Requests | Exponential backoff | Not stated by xAI on the page checked | Rate limits |
Batch and asynchronous limits
Batch interfaces usually have their own, larger limits and are a common way to move bulk work off the real-time quota.
- Anthropic: Message Batches limits are shared across models: Start 1,000 RPM, 200,000 requests in the processing queue; Build 2,000 RPM, 300,000; Scale 4,000 RPM, 500,000; 100,000 requests per batch at every tier. Batches are not subject to ramp limits.
- Cohere: EmbedJob is limited to 5 requests a minute on trial keys and 50 on production keys.
- DeepSeek: no batch limits stated on the rate limit page.
- Google: 100 concurrent batch requests, 2 GB input file, 20 GB file storage, and per-model enqueued token limits by tier.
- Meta: background responses capped at 600 submissions a minute per team.
- Mistral AI: no batch limits stated on the tier page.
- OpenAI: batch queue limits are calculated from total input tokens queued per model; pending jobs count until they complete.
- xAI: no batch limits stated on the rate limit page; the pricing page lists a 20 percent batch discount for grok-4.7.
How to raise your limits
- Measure first. Log the rate limit headers each provider returns (where documented) for a week so you know which dimension you actually hit: requests, input tokens, output tokens, concurrency or daily quota.
- Use the cheap levers. Prompt caching raises effective input throughput on Anthropic; batch interfaces move bulk jobs off real-time limits at Anthropic, Google and OpenAI; DeepSeek's off-peak hours and xAI's batch discount reduce cost, not limits.
- Earn the next tier. Spend-based tiers (Google, OpenAI, xAI) move automatically once thresholds and any waiting periods are met. Google's Tier 2 and Tier 3 also need 3 and 30 days from first payment, so plan launches around them.
- Ask. Anthropic (sales via Console above Scale), Cohere (support), DeepSeek (expansion form), Google (increase form), Mistral AI (support) and xAI (Console request, or sales for Enterprise) each document a request route.
- Split traffic deliberately. Because scope differs, extra projects or keys do not add capacity on Google (project), DeepSeek (account) or Meta and xAI (team). On OpenAI, project limits sit inside the organisation limit. See the AI Agent Rate Limits guide for quota design inside your own platform.
Where identity and security fit
Rate limits are also a security control. Because limits sit on the organisation, project, team or account, one leaked key can exhaust capacity for every application sharing that scope, and a sudden jump to tier limits can be a sign that a stolen key is in use. The LLMjacking Guide covers that pattern, and the key guides for Anthropic API Keys Security Guide, Google API Keys Security Guide and OpenAI API Keys Security Guide cover scoping and rotation. Putting providers behind a gateway lets you enforce per-application quotas below the provider's limits; see the AI Gateway and LLM Gateway Guide.
Common failure patterns
- Treating every 429 as transient: Anthropic's spend cap error and OpenAI's quota and billing errors will not clear with retries; Anthropic's cap resumes at 00:00 UTC on the first of the next month.
- Adding keys to add capacity: on Google, Meta, xAI and DeepSeek the limit is on the project, team or account, so new keys share the same bucket.
- Sizing on max_tokens: Anthropic counts generated tokens for OTPM, not
max_tokens, while xAI counts cached prompt tokens in TPM; copying one provider's sizing to another misjudges headroom. - Launching on a fresh account: Google's waiting periods and Anthropic's Evaluation tier mean a new organisation can start with low limits.
- Retry storms: failed requests still count on OpenAI, and client SDK retries can double your own retry logic. Disable one layer.
Practitioner checklist
- Record, per provider, the limit units, scope and your current tier, and keep it next to your capacity plan.
- Read the provider's console limits page for each model in production; public docs rarely show every model.
- Log rate limit headers where the provider documents them and alert at 80 percent of any dimension.
- Honour
Retry-Afteras a minimum, add jitter, cap total retry time, and never retry spend or billing errors. - Route bulk and offline jobs to batch interfaces and use prompt caching where it reduces counted input.
- Ramp new traffic gradually, within Anthropic's ramp and OpenAI's growth guidance.
- Give each application its own project or workspace where the provider supports it, and set internal quotas through a gateway.
- Treat an unexplained jump to tier limits as a possible key compromise and check the key's usage logs.
Standards and references
- Anthropic: Rate limits (accessed 2026-11-15)
- Cohere: Rate Limits (accessed 2026-11-15)
- DeepSeek: Rate Limit (accessed 2026-11-15)
- Google: Gemini API rate limits (accessed 2026-11-15)
- Meta: Model API pricing and rate limits (accessed 2026-11-15)
- Mistral AI: Rate limit and usage tiers (accessed 2026-11-15)
- OpenAI: Rate limits (accessed 2026-11-15)
- xAI: Rate limits (accessed 2026-11-15)
Related NHI Mgmt Group resources: LLM API 401, 403 and 429 Errors Guide, LLM API Free Tiers and Credits Compared, LLM Pricing Tracker, Frontier AI Models Compared.