Home / LLM rate limits
LLM API rate limits, compared
Source-reviewed subset: 18 providers. Exact model, tier and scope are retained below. This is not a full provider catalog. Re-check the linked provider docs and your account limits before capacity planning.
Limits are account-scoped and providers change them; re-check the linked docs before capacity planning.
Concurrency caps are not guaranteed throughput and are never shown here as RPM equivalents.
Prompt-cache accounting is model-specific; see each provider's notes instead of a blanket rule.
18 of 18 providers
Quick comparison
| Provider | Free tier | Limit style | Checked |
|---|---|---|---|
| OpenAI | Not reviewed | Build, Standard: Astra/Sol/Terra: 5,000 RPM; 1,000,000 TPM | 2026-10-26 |
| Anthropic | Not reviewed | Start: Claude Opus 5 / Sonnet 5 / Haiku 4.5: 1,000 RPM; 2,000,000 input TPM; 400,000 output TPM for each listed model | 2026-10-26 |
| Google Gemini API | Not reviewed | Interactive requests, each project/model: RPM, input TPM and RPD vary by model and account; no universal numeric ceiling shown here | 2026-10-26 |
| Groq | Not reviewed | Organization/model-specific: Numeric figures held: primary rate-limit and model pages disagree for the Developer plan | 2026-10-26 |
| Mistral | Free variants | Organization/model limits: Per-model requests/second and tokens/minute; numeric ceilings are account-specific | 2026-10-26 |
| xAI | Not reviewed | grok-4.7, Tier 0: 150 requests/second; 50M TPM | 2026-10-26 |
| Cohere | Trial | Command A / R+ / R / R7B, evaluation key: 20 requests/minute per model | 2026-10-26 |
| Perplexity | Not reviewed | Agent API, Tier 0: 1 query/second shared across models; per-model 5M TPM and 300M TPH | 2026-10-26 |
| Cerebras | Trial | Organization/model-specific, dual token buckets: Uncached TPM and total TPM are independent. Default total ceiling is 3× uncached, with model-specific exceptions | 2026-10-26 |
| SambaNova | Free variants | SambaCloud Free: Model-specific RPM, RPD and TPD; exact table tier assignment not inferred from flattened tabs | 2026-10-26 |
| Fireworks AI | Not reviewed | Serverless Small, e.g. OpenAI GPT OSS 120B: 216M total-prompt TPM; 54M uncached-prompt TPM; 2.16M generated TPM | 2026-10-26 |
| DeepSeek | Not reviewed | deepseek-v4-pro: 500 concurrent requests | 2026-10-26 |
| DeepInfra | Not reviewed | Account default, each model: 200 concurrent requests per model | 2026-10-26 |
| OpenRouter | Free variants | :free variants, lower credit tier: 20 RPM; 50 requests/day | 2026-10-26 |
| Together AI | Not reviewed | Serverless: No universal numeric rate published on this page; best-effort, model/demand-specific throttling | 2026-10-26 |
| Cloudflare Workers AI | Not reviewed | Text Generation task default: 300 requests/minute | 2026-10-26 |
| Amazon Bedrock | Not reviewed | Account/model/endpoint quotas: No single provider-wide numeric ceiling; runtime and mantle have separate per-model allocations | 2026-10-26 |
| Azure OpenAI in Foundry | Not reviewed | gpt-4.1, Tier 1, GlobalStandard: 1,000 RPM; 1M TPM | 2026-10-26 |
Provider detail
OpenAI
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Build, Standard: Astra/Sol/Terra | Requirement: $5 total credit purchases; approved usage limit $500/month | Limits: 5,000 RPM; 1,000,000 TPM |
| Build, Standard: Luna | Requirement: $5 total credit purchases; approved usage limit $500/month | Limits: 5,000 RPM; 2,000,000 TPM |
| Launch, Standard: Astra/Sol/Terra | Requirement: $100 total credit purchases; approved usage limit $5,000/month | Limits: 10,000 RPM; 4,000,000 TPM |
| Launch, Standard: Luna | Requirement: $100 total credit purchases; approved usage limit $5,000/month | Limits: 10,000 RPM; 10,000,000 TPM |
| Grow, Standard: Astra/Sol/Terra | Requirement: $500 total credit purchases; approved usage limit $200,000/month | Limits: 15,000 RPM; 40,000,000 TPM |
| Grow, Standard: Luna | Requirement: $500 total credit purchases; approved usage limit $200,000/month | Limits: 30,000 RPM; 180,000,000 TPM |
- Published Standard table only, not Ultrafast or all models. Limits vary by model and can be shared by model families. Organization and project scope, not per user.
- Monthly approved usage limits, configured spend caps, and rate limits are distinct. A Free usage-tier label does not establish free API credit.
- Long-context requests can have separate limits. Verify account limits in the provider console.
Provider documentation Checked 2026-10-26
Anthropic
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Start: Claude Opus 5 / Sonnet 5 / Haiku 4.5 | Requirement: Organization Start tier; Evaluation may start lower | Limits: 1,000 RPM; 2,000,000 input TPM; 400,000 output TPM for each listed model |
| Build: Claude Opus 5 / Sonnet 5 / Haiku 4.5 | Requirement: Organization Build tier | Limits: 5,000 RPM; 5,000,000 input TPM; 1,000,000 output TPM for each listed model |
| Scale: Claude Opus 5 / Sonnet 5 / Haiku 4.5 | Requirement: Organization Scale tier | Limits: 10,000 RPM; 10,000,000 input TPM; 2,000,000 output TPM for each listed model |
- Selected standard Messages API models only. Separate Fable, older-model shared buckets, batch and fast-mode limits are not represented by these figures.
- Organization scope, with workspace configuration possible. Short bursts and acceleration limits can trigger 429. Maximum allowed rates are not guaranteed capacity.
- For these selected models cache reads do not count toward input TPM; cache creation does. Haiku 3.5 is an exception and is not included here. AWS Marketplace billing and tier-management rules differ.
Provider documentation Checked 2026-10-26
Google Gemini API
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Interactive requests, each project/model | Requirement: Project usage tier; exact active limits in AI Studio | Limits: RPM, input TPM and RPD vary by model and account; no universal numeric ceiling shown here |
| Batch API | Requirement: Separate from interactive limits | Limits: 100 concurrent batch requests; model/tier-specific enqueued-token ceilings |
- Project scope, not API-key scope. RPD resets at midnight Pacific. Specified limits are not guaranteed capacity.
- Spend-based rolling-window limits and billing-tier caps are separate from RPM/TPM. This view does not infer free model availability from the Free tier label.
- The source also documents batch file/storage caps; these are not request throughput.
Provider documentation Checked 2026-10-26
Groq
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Organization/model-specific | Requirement: Check current account Limits page | Limits: Numeric figures held: primary rate-limit and model pages disagree for the Developer plan |
- For openai/gpt-oss-120b and 20b, the rate-limit page shows 30 RPM / 8K TPM while the model page labels Developer plan as 1K RPM / 250K TPM. No plan assignment is inferred from this conflict.
- Organization scope. RPM, RPD, TPM and TPD can bind independently; some organizations also have input/output TPM caps. Source says cached tokens do not count.
- Conflicting model-page source: https://console.groq.com/docs/models. Use the live account limit as authority before planning.
Provider documentation Checked 2026-10-26
Mistral
Free variants available| Tier / model | Requirement | Published limits |
|---|---|---|
| Organization/model limits | Requirement: Plan and current Admin Panel API Limits | Limits: Per-model requests/second and tokens/minute; numeric ceilings are account-specific |
| Free mode | Requirement: Included monthly usage within current Limits page | Limits: No universal numeric free-mode throughput stated on this page |
- Source explicitly describes Free mode and included monthly usage, not unlimited access to every model. Pay-as-you-go extends usage through prepaid credits.
- Requests/second and tokens/minute are independent. Monthly organization/workspace spend caps are not rate limits.
- Known limitations page says batch processing does not count against real-time limits: https://docs.mistral.ai/resources/known-limitations.
Provider documentation Checked 2026-10-26
xAI
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| grok-4.7, Tier 0 | Requirement: Team default tier; $0 cumulative-spend threshold, not free credit | Limits: 150 requests/second; 50M TPM |
| grok-4.7, Tier 4 | Requirement: $5,000 qualifying cumulative API spend since 1 Jan 2026 | Limits: 500 requests/second; 100M TPM |
- Selected exact model and two tier endpoints only, not all Grok models. Team per-model limits; RPS is not RPM. Check personalized limits in Console.
- TPM includes prompt, completion, reasoning and cached prompt tokens. Other providers may account for cache differently.
- Tiers depend on qualifying credit purchases/fulfilled invoices and do not downgrade; a spend threshold is not a usage allowance.
Provider documentation Checked 2026-10-26
Cohere
Time-boxed trial| Tier / model | Requirement | Published limits |
|---|---|---|
| Command A / R+ / R / R7B, evaluation key | Requirement: Free evaluation key; 1,000 API calls/month | Limits: 20 requests/minute per model |
| Command A / R+ / R / R7B, production key | Requirement: Paid production key | Limits: 500 requests/minute per model |
| Command A+ / A Reasoning / A Translate / A Vision | Requirement: Trial-like limits can apply even with production key | Limits: 20 requests/minute evaluation; production access requires provider confirmation |
- Evaluation is not unlimited production access. Newer Chat variants may retain the 1,000-call monthly limit on production keys.
- Chat subset only. Embed uses inputs/minute, not requests/minute; Rerank and other endpoints have separate limits.
- This view does not infer account-wide quota sharing beyond the per-model/key distinctions published here.
Provider documentation Checked 2026-10-26
Perplexity
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Agent API, Tier 0 | Requirement: New organization / tier 0 | Limits: 1 query/second shared across models; per-model 5M TPM and 300M TPH |
| Agent API, Tier 5 | Requirement: $5,000+ lifetime credits purchased | Limits: 33 queries/second shared across models; per-model 16M TPM and 960M TPH |
- Agent API scope only, not a Sonar or Search API blanket rate. Request limits are organization-wide across models/keys; token budgets are organization-and-model scoped.
- Poll, cancel, resume and file-download calls also consume the shared request budget. Token and request limits bind independently.
- Search API uses query units/second, not Agent QPS. Tier 0 is a spending tier, not evidence of free service.
Provider documentation Checked 2026-10-26
Cerebras
Time-boxed trial| Tier / model | Requirement | Published limits |
|---|---|---|
| Organization/model-specific, dual token buckets | Requirement: Free Trial, Developer or Enterprise; exact limits in Console | Limits: Uncached TPM and total TPM are independent. Default total ceiling is 3× uncached, with model-specific exceptions |
- Cached tokens remain in total TPM but not uncached TPM. qwen-3.8-27b documents a temporary 5× total-token exception. No generic cache multiplier promises throughput.
- Published tier tables are retained in the research capture, but numeric plan assignment is not inferred from flattened tab content in this preview.
- Requests and token limits can both bind. Estimated input plus upper-bound output can cause pre-request throttling; actual usage updates the quota.
Provider documentation Checked 2026-10-26
SambaNova
Free variants available| Tier / model | Requirement | Published limits |
|---|---|---|
| SambaCloud Free | Requirement: No payment method linked | Limits: Model-specific RPM, RPD and TPD; exact table tier assignment not inferred from flattened tabs |
| SambaCloud Developer | Requirement: Payment method linked | Limits: 20M tokens/day across all models, alongside model-specific request limits |
- Free tier is explicitly documented, not inferred from a zero-spend tier. Exact model availability and account limits can differ.
- SambaStack deployments instead allow administrators to apply optional group limits. Do not apply SambaCloud defaults to SambaStack.
- Preview models are evaluation-only and can be removed at short notice.
Provider documentation Checked 2026-10-26
Fireworks AI
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Serverless Small, e.g. OpenAI GPT OSS 120B | Requirement: Maximum base ceilings at spending Tier 3+; current adaptive limit can be lower | Limits: 216M total-prompt TPM; 54M uncached-prompt TPM; 2.16M generated TPM |
| Serverless Large, e.g. Kimi K3 | Requirement: Maximum base ceilings at spending Tier 3+; current adaptive limit can be lower | Limits: 21.6M total-prompt TPM; 5.4M uncached-prompt TPM; 216K generated TPM |
- These are maximum base ceilings, not default/current account throughput. The three metrics adapt separately with recent usage. Lower spending tiers have lower ceilings.
- Total model parameters determine size tier, not active parameters. Unknown parameter count uses Large. Fast/Priority/US-only variants share the base model tier.
- Eligible Elastic Burst can temporarily exceed base ceilings; spare capacity is not guaranteed. Account-specific dashboard limits govern. No RPM equivalent invented.
Provider documentation Checked 2026-10-26
DeepSeek
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| deepseek-v4-pro | Requirement: Account default | Limits: 500 concurrent requests |
| deepseek-v4-flash | Requirement: Account default | Limits: 2,500 concurrent requests |
| deepseek-v4-flash-vision-exp | Requirement: Account default | Limits: 2,500 concurrent requests |
- A connection counts from request send until the model response completes. Account-level limit across API keys.
- Regular users share the account cap across user_id values. Expanded-quota accounts also have per-user_id caps. Capacity expansion may be requested at no additional cost.
- The cited page specifies concurrency, not a guaranteed requests-per-minute rate.
Provider documentation Checked 2026-10-26
DeepInfra
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Account default, each model | Requirement: Account | Limits: 200 concurrent requests per model |
- Concurrency is not RPM. Throughput depends on request duration. Two distinct models can each use 200 concurrent requests.
- Limit increases can be requested through Dashboard > Account.
Provider documentation Checked 2026-10-26
OpenRouter
Free variants available| Tier / model | Requirement | Published limits |
|---|---|---|
| :free variants, lower credit tier | Requirement: Less than 10 lifetime credits purchased (published table) | Limits: 20 RPM; 50 requests/day |
| :free variants, higher credit tier | Requirement: At least 10 lifetime credits purchased (published table) | Limits: 20 RPM; 1,000 requests/day |
- Free variants only, not a paid-model default. Docs state the higher ceiling starts one credit below the table threshold to absorb rounding/top-up fees (currently 9 credits). Check account ceiling.
- Some endpoints/accounts and BYOK requests are exempt. Upstream provider throttling and DDoS protection can still apply.
- Credit/budget exhaustion (402) is distinct from request throttling (429).
Provider documentation Checked 2026-10-26
Together AI
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Serverless | Requirement: Serverless inference | Limits: No universal numeric rate published on this page; best-effort, model/demand-specific throttling |
| Provisioned throughput | Requirement: Contract for selected model or model family | Limits: Committed capacity under the agreed SLA |
- High volume or bursts can receive 429 or 503. Model-specific restrictions may apply at high demand.
- Existing Enterprise/Scale purchased limits remain in force until contract expiry. This is not an unlimited-throughput promise.
Provider documentation Checked 2026-10-26
Cloudflare Workers AI
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Text Generation task default | Requirement: Workers AI platform default; model exceptions apply | Limits: 300 requests/minute |
| @cf/qwen/qwen1.5-14b-chat-awq | Requirement: Exact named model exception | Limits: 150 requests/minute |
| @cf/tinyllama/tinyllama-1.1b-chat-v1.0 | Requirement: Exact named model exception | Limits: 720 requests/minute |
- Task defaults are not universal model limits. Beta models may have lower limits. Local Wrangler inference counts toward usage.
- Speech, embeddings, image and OCR tasks have different units/limits. This source alone does not verify free usage allocation or quota scope.
Provider documentation Checked 2026-10-26
Amazon Bedrock
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| Account/model/endpoint quotas | Requirement: Current Service Quotas values, regional/account conditions | Limits: No single provider-wide numeric ceiling; runtime and mantle have separate per-model allocations |
- Model inference uses token quotas with model-specific accounting. Embedding models are an exception, throttled by RPM.
- Account age, usage/payment history and regional factors can produce lower quotas than published defaults. Increases are not automatic.
- The same underlying model on bedrock-runtime and bedrock-mantle does not share a single interchangeable quota.
Provider documentation Checked 2026-10-26
Azure OpenAI in Foundry
Free access not reviewed| Tier / model | Requirement | Published limits |
|---|---|---|
| gpt-4.1, Tier 1, GlobalStandard | Requirement: Exact model, deployment type and subscription quota tier | Limits: 1,000 RPM; 1M TPM |
| gpt-4.1, Tier 1, DataZoneStandard | Requirement: Exact model, deployment type and subscription quota tier | Limits: 300 RPM; 300K TPM |
- Subscription scope, not tenant scope. Subscription-level migration began with Realtime Translate/Whisper; do not generalize rollout completion to every model.
- Global Standard same-model/version pools span subscription regions; Data Zone Standard pools are per data zone where this management applies.
- Tier 1 selected rows only, not all Azure models or tenants. Some other table rows use a 10-second request window; they are not silently read as RPM.
Provider documentation Checked 2026-10-26
Methodology
- This subset was checked against the linked provider documentation on 2026-10-26. Each card retains the exact model/tier scope; numbers are not provider-wide guarantees.
- A page without a universal numeric rate is not evidence of unlimited capacity. Contract-specific and account-specific limits can differ.
- Free access is marked not reviewed unless this subset directly verifies a free model variant. A usage-tier name or spending cap is not free credit.
- RPM = requests per minute; TPM = tokens per minute; TPD/RPD = per day; RPS = requests per second. Cache-read accounting is model-specific and is not inferred for this subset.
See also: LLM pricing comparison for per-token costs across these providers.