API Rate Limits Compared: Every Major LLM Provider (April 2026)

Second edition. The first (March 2026) covered 15 providers. This update adds Cerebras and SambaNova, covers new flagship models (GPT-5.5, Claude Opus 4.7, DeepSeek V4, Gemini 3.1 Pro), and reflects structural changes at several providers since March.

What changed since March
Free Tier — RPM & RPD
Free Tier — TPM & TPD
Entry Paid Tier — RPM & RPD
Entry Paid Tier — TPM
Scaled Tier — RPM & TPM
New Providers — Cerebras & SambaNova
Audio Models — ASH & ASD
Cloud Aggregators — Azure AI & AWS Bedrock
More Providers — Perplexity, Alibaba, Moonshot
Provider Notes
Tips for managing rate limits

Last updated: April 25, 2026.

What changed since March

New models in rate limit tables:

OpenAI: GPT-5.5 (Apr 24) and GPT-5.4 mini/nano (Mar 17) joined the lineup. GPT-5.5 shares rate limits with GPT-5.4. Free tier still excluded from all GPT-5.x models.
Anthropic: Claude Opus 4.7 (Apr 16) shares the Opus 4.x rate limit pool with Opus 4.6. No separate bucket.
DeepSeek: V4-Flash and V4-Pro (Apr 24) replaced V3.2 as the current models. Rate limits remain fully dynamic (no published caps).
Google: Gemini 3.1 Pro and 3.1 Flash-Lite are in preview with conservative limits. Gemini 2.0 Flash and Flash-Lite deprecated; shutting down June 1, 2026.

Structural changes:

Google removed static rate limit tables from public docs. Limits are now only visible in AI Studio (aistudio.google.com/rate-limit). The numbers below for 2.5-series models are the last confirmed published figures.
Together AI switched to fully dynamic rate limits in January 2026. No fixed tiers; limits grow with sustained usage.
Groq now hosts GPT-OSS (20B/120B) and Qwen3-32B on its free tier.

New providers added: Cerebras (published free tier, fastest inference) and SambaNova (published free tier, best TTFT).

What are rate limits?

Rate limits cap how much you can use an API within a given time window. Providers enforce them to manage capacity and ensure fair access. Exceeding a limit returns a 429 Too Many Requests error.

Most providers use the token bucket algorithm: capacity refills continuously up to your maximum, rather than resetting at fixed intervals. A 60 RPM limit means roughly 1 request per second with steady refill, not 60 requests then a hard stop.

The metrics

RPM (Requests per minute): API calls per minute, regardless of size. A 10-token request and a 100K-token request both count as one.
RPD (Requests per day): Daily cap on total API calls. Some providers use this instead of (or alongside) RPM, especially on free tiers.
TPM (Tokens per minute): Total tokens (input + output) processed per minute. Usually the binding constraint in production.
TPD (Tokens per day): Daily token cap. More common on free tiers.
ITPM/OTPM (Input/Output tokens per minute): Anthropic separates input and output token limits, giving finer control. Cached input tokens don’t count toward ITPM on current Claude 4.x models.
ASH/ASD (Audio seconds per hour/day): For speech models like Whisper.

RPM limits set your concurrency ceiling. TPM limits set your throughput. For batch workloads, RPD and TPD matter more. For real-time apps, RPM and TPM hit first.

Free Tier — Requests per Minute (RPM) & Requests per Day (RPD)

Metric	Google			Groq		Cerebras	SambaNova	Fireworks AI	Cohere
Metric	2.5 Pro	2.5 Flash	2.5 Flash-Lite	Llama 4 Scout	Qwen3 32B	Llama 3.3 70B	Llama 3.1 405B	OS models	Command R+
RPM	5	10	15	30	60	30	10–30	10	20
RPD	50–100	250–1,500	1,000	1,000	1,000	—	—	—	~33*

*Cohere trial keys get 1,000 calls/month across all endpoints, roughly ~33/day.

Note: OpenAI, Anthropic, xAI, DeepSeek, and Mistral have no free API tier for current flagship models. OpenAI’s free tier does not support any GPT-5.x model. Anthropic’s lowest entry is $5 (Tier 1). xAI and Mistral gate numerical limits behind their respective consoles.

Change from March: Google’s free tier RPD dropped 50–80% in late 2025. Gemini 2.5 Pro went from ~100–200 RPD to 50–100. Groq added GPT-OSS (20B/120B) and Qwen3-32B to its free tier.

Free Tier — Tokens per Minute (TPM) & Tokens per Day (TPD)

Metric	Google			Groq		Cerebras	SambaNova
Metric	2.5 Pro	2.5 Flash	2.5 Flash-Lite	Llama 4 Scout	Qwen3 32B	Llama 3.3 70B	Llama 3.1 405B
TPM	250,000	250,000	250,000	30,000	6,000	60,000	—
TPD	—	—	—	500,000	500,000	1,000,000	—

Standout: Google still leads with 250K TPM on the free tier, but the RPM/RPD caps (5–15 RPM) mean it’s best for fewer, larger requests. Cerebras offers 60K TPM with 1M TPD and ~2,100 tokens/second inference speed on Llama 3.3 70B, making it the fastest free option by throughput. SambaNova publishes RPM but not TPM/TPD specifics.

Entry Paid Tier — Requests per Minute (RPM) & Requests per Day (RPD)

The tier most developers start with. OpenAI Tier 1 ($5), Anthropic Tier 1 ($5), Google Tier 1 (pay-as-you-go), and others at their lowest paid level.

Metric	OpenAI		Anthropic		Google			xAI	DeepSeek	Fireworks AI	Cohere
Metric	GPT-5.5	GPT-5.4 nano	Sonnet 4.x	Haiku 4.5	2.5 Pro	2.5 Flash	2.5 Flash-Lite	Grok 4.20	V4 Flash	OS models	Command R+
RPM	500	500	50	50	150	300	300	Console†	Dynamic‡	≤6,000	500
RPD	—	—	—	—	1,000	1,500	1,500	—	—	—	—

†xAI publishes tier thresholds ($0/$50/$250/$1K/$5K) but numerical RPM/TPM are only visible in the xAI Console after login. ‡DeepSeek uses fully dynamic concurrency limits based on server load. No fixed RPM/TPM published.

Change from March: Mistral also moved to console-only limits. Actual numbers require login at admin.mistral.ai/plateforme/limits. Both xAI and Mistral are excluded from RPM columns where no public figure exists.

Standout: Fireworks can spike to 6,000 RPM but it’s a dynamic ceiling, not guaranteed (soft limit starts at ~1 RPS and doubles hourly). Anthropic’s 50 RPM at Tier 1 is the lowest here, but jumps 20x to 1,000 RPM at Tier 2 ($40).

Entry Paid Tier — Tokens per Minute (TPM)

Metric	OpenAI		Anthropic		Google			DeepSeek
Metric	GPT-5.5	GPT-5.4 nano	Sonnet 4.x	Haiku 4.5	2.5 Pro	2.5 Flash	2.5 Flash-Lite	V4 Flash
TPM	500K	200K	30K in / 8K out	50K in / 10K out	1M	2M	2M	Dynamic

Standout: Google 2.5 Flash and Flash-Lite lead at 2M TPM. GPT-5.5 gets 500K TPM at Tier 1 (same as GPT-5.4). Anthropic looks low on paper (30K ITPM for Sonnet), but cached input tokens don’t count toward the limit. With 80% cache hit rate, effective throughput is 5x higher (150K+ effective ITPM).

Scaled Tier — RPM & TPM at Higher Spend

For teams past entry-level. OpenAI Tier 3 ($100+), Anthropic Tier 3 ($200+), Google Tier 2 ($250+).

Metric	OpenAI (Tier 3)		Anthropic (Tier 3)		Google (Tier 2)
Metric	GPT-5.5	GPT-5.4 nano	Opus 4.x	Haiku 4.5	2.5 Pro	2.5 Flash
RPM	5,000	5,000	2,000	2,000	1,000	2,000
TPM	2M	4M	800K in / 160K out	1M in / 200K out	2M	4M

At the highest standard tiers (OpenAI Tier 5 at $1,000+, Anthropic Tier 4 at $400+):

Metric	OpenAI (Tier 5)		Anthropic (Tier 4)
Metric	GPT-5.5	GPT-5.4 mini	Opus 4.x	Haiku 4.5
RPM	15,000	30,000	4,000	4,000
TPM	40M	180M	2M in / 400K out	4M in / 800K out

Standout: OpenAI’s Tier 5 numbers are staggering. GPT-5.4 mini at 180M TPM and 30K RPM is built for high-volume classification and routing. GPT-5.5 gets 15K RPM and 40M TPM. Anthropic’s Tier 4 caps at 4K RPM but cached tokens are free, so effective throughput can be 5x+ higher with good cache hit rates.

Change from March: GPT-5.5 now matches GPT-5.4 at every tier (identical rate limit profiles). The old GPT-4.1 nano Tier 5 figure of 150M TPM has been overtaken by GPT-5.4 mini at 180M TPM.

New Providers — Cerebras & SambaNova

Both specialize in custom silicon for inference speed. Neither existed in the March edition.

Cerebras

Hardware: Wafer-Scale Engine (WSE-3). Fastest published inference: ~2,100 tokens/second on Llama 3.3 70B.

Tier	RPM	TPM	TPD
Free	30	60,000	1,000,000
Paid	Higher (contact sales)	Higher	Higher

Hosts Llama 3.3 70B, Llama 3.1 8B, and other open models. The speed advantage is real: tasks that take 30 seconds on GPU-based providers finish in under 5 seconds. Paid tier limits are not publicly documented.

SambaNova

Hardware: Custom RDU (Reconfigurable Dataflow Unit). Best time-to-first-token (TTFT): ~0.2 seconds.

Tier	RPM	Notes
Free	10–30 (varies by model)	Hosts up to Llama 3.1 405B for free

SambaNova is notable for offering Llama 3.1 405B on the free tier. Most other free-tier providers cap out at 70B-class models. TPM and TPD limits are not publicly documented.

Audio Models — ASH & ASD

Groq and Fireworks both publish audio-specific limits.

Metric	Groq (Free Tier)		Fireworks AI
Metric	Whisper Large v3	Whisper Large v3 Turbo	Whisper v3-large	Whisper v3-turbo
RPM	20	20	—	—
RPD	2,000	2,000	—	—
ASH	7,200	7,200	—	—
ASD	28,800	28,800	—	—
Audio min/min	—	—	200	400

Groq: 7,200 ASH = 2 hours of audio per hour of wall time. 28,800 ASD = 8 hours per day. Plenty for a podcast transcription pipeline on the free tier.

Fireworks: 200 min/min for Whisper v3-large, 400 min/min for v3-turbo. Concurrent streaming capped at 10 connections.

Cloud Aggregators — Azure AI & AWS Bedrock

Both use configurable per-deployment limits, not fixed tiers. The shared ratio is 6 RPM per 1,000 TPM.

Azure AI (Microsoft)

Deployment Type	How Limits Work
Pay-as-you-go (Standard)	TPM quota per model per region, auto-scales with usage
Provisioned (PTU)	Reserved throughput units, no per-request limits
Global/Data Zone	Higher default quotas, multi-region routing

Multi-model: admins can select GPT-5.5, GPT-5.4, Claude Opus/Sonnet 4.6, or Gemini 3.1 Pro. Azure AI Quotas & Limits

AWS Bedrock

Hosts Claude (Opus 4.7, Sonnet 4.6, Haiku 4.5), Llama 4, Mistral, and Nova models with per-model, per-region quotas. Default quotas vary by model and can be increased via AWS Service Quotas console. Provisioned Throughput deployments remove rate limits entirely. AWS Bedrock Quotas

More Providers — Perplexity, Alibaba (Qwen), Moonshot (Kimi)

Metric	Perplexity			Alibaba (Qwen)			Moonshot
Metric	Sonar Pro	Sonar	Deep Research	Qwen3 Max	Qwen3.5 Plus	Qwen3.5 Flash	Kimi K2.5
RPM (T0)	50	50	5	—			3
RPM (T1)	150	150	10	600	15,000	15,000	200
RPM (T3+)	1,000	1,000	40	600	30,000	30,000	5,000
TPM	—	—	—	1M	5M	10M	—
TPD (T0)	—	—	—	—	—	—	1.5M

Perplexity tiers: T0 = new account, T1 = $50+, T3 = $500+. Alibaba limits shown for international (Singapore) deployment; Beijing deployment is higher (Qwen3 Max jumps to 30K RPM). Moonshot T1 = $10+ cumulative recharge; T0 has 1.5M TPD cap, T1+ is unlimited.

Change from March: Perplexity added Sonar Deep Research (very low limits: 5–100 RPM depending on tier) and an Agent API (50–2,000 RPM across tiers). Moonshot expanded from 2 tiers (T0–T1) to 6 tiers (T0–T5), with T5 reaching 10K RPM and 5M TPM at $3,000+ cumulative recharge.

Standout: Alibaba’s Qwen3.5 Flash at 15K RPM and 10M TPM remains the highest throughput in this entire post. Moonshot’s new tier structure is competitive at the high end (T5: 10K RPM, 1,000 concurrent connections).

Provider Notes

OpenAI: GPT-5.5 and GPT-5.4 share identical rate limit profiles across all tiers. GPT-5.4 mini at Tier 5 (180M TPM, 30K RPM) is the throughput king. Free tier excluded from all GPT-5.x models. Batch API offers 50% discount with higher queue limits. Assistants API deprecated August 2026; replaced by Responses API. Docs
Anthropic: Opus 4.7 shares a combined rate limit pool with all Opus 4.x variants. You cannot independently max out Opus 4.7 and 4.6 simultaneously. Cached input tokens don’t count toward ITPM on current 4.x/4.5 models. No free tier; lowest entry is $5. Fast mode on Opus 4.6 draws from a separate dedicated pool. Docs
Google: Removed static rate limit tables from public docs in Q1 2026. Actual limits now only visible in AI Studio dashboard. Gemini 3.1 Pro and 3.1 Flash-Lite are in preview with conservative limits. Gemini 2.0 Flash/Flash-Lite shut down June 1, 2026. Docs
Groq: Free tier publishes exact numbers (30 RPM, 6K–30K TPM depending on model). Daily token budgets (TPD) alongside TPM. Cached tokens don’t count. Now hosts GPT-OSS and Qwen3-32B. Runs on custom LPU hardware. Docs
xAI: 5-tier structure ($0/$50/$250/$1K/$5K thresholds) but numerical RPM/TPM only visible in xAI Console. Docs
Mistral: European data residency (GDPR). 5-tier structure (Free through Tier 4 at $500+). Limits enforced per RPS (not RPM), TPM, and tokens/month. Actual numbers require login at admin.mistral.ai. New model: Mistral Small 4 (v26.03). Docs
DeepSeek: Fully dynamic concurrency limits based on server load. No published RPM/TPM. V4-Flash and V4-Pro are the current models (V3.2 aliases deprecated July 24, 2026). V4-Flash pricing: $0.14/M input (cache miss), $0.028/M (cache hit), $0.28/M output. Docs
Cerebras: Fastest inference (~2,100 tok/s). Free tier: 30 RPM, 60K TPM, 1M TPD. Paid limits not publicly documented. Custom WSE-3 silicon.
SambaNova: Best TTFT (~0.2s). Free tier: 10–30 RPM depending on model size. Hosts up to Llama 3.1 405B for free. Custom RDU hardware.
Together AI: Dynamic rate limits since January 2026. No fixed tiers or published numbers. Limits grow with sustained usage and are returned in API response headers. Docs
Fireworks AI: Dynamic ceiling up to 6,000 RPM (soft limit starts ~1 RPS, doubles hourly). On-demand GPU deployments remove limits. Spending tier caps: $50–$50K/month by tier. Docs
Cohere: Chat endpoint: 20 RPM (trial), 500 RPM (production). Rerank: 1,000 RPM production. Embed: 2,000 inputs/min. Trial keys capped at 1,000 calls/month. No TPM limits published. Docs
Perplexity: 6 tiers (T0–T5) based on cumulative spend. Leaky bucket algorithm. Deep Research model has very low limits (5–100 RPM). New Agent API: 50–2,000 RPM by tier. Docs
Alibaba (Qwen): Region-specific limits. Beijing deployment is more generous than Singapore/Global. Qwen3.5 Flash at 30K RPM / 10M TPM (Beijing) is the highest throughput of any provider listed. Docs
Moonshot (Kimi): Expanded to 6 tiers (T0–T5). T0 ($1 recharge): 3 RPM, 1 concurrent, 1.5M TPD. T5 ($3,000): 10K RPM, 1,000 concurrent, 5M TPM, unlimited TPD. Automatic 75% caching discount. Docs
AWS Bedrock: Hosts Claude, Llama, Mistral, Nova with per-model per-region quotas. 6 RPM per 1K TPM ratio. Provisioned Throughput removes limits. Docs
NVIDIA NIM: Hosted API is for prototyping (40 RPM). Self-hosted NIM containers have no limits. Docs

Tips for managing rate limits

For a deep dive on token pricing, cost optimization strategies, caching mechanics, model routing architectures, and production cost modeling, see LLM Token Costs and Efficiency. This section focuses on rate limit management specifically.

Handling 429s

Use exponential backoff when you hit 429s. The Anthropic SDK, OpenAI SDK, and most third-party clients handle this automatically. Don’t build your own retry loop unless you need custom jitter or circuit-breaking behavior.
Check response headers before you hit the wall. Together AI and Fireworks return your current rate limit state in every API response. DeepSeek adjusts concurrency limits dynamically based on server load. Read these headers in production to implement proactive throttling rather than reactive retries.

Caching for throughput (not just cost)

Prompt caching has a rate limit benefit that is separate from its cost benefit. The cost savings are covered in the token costs post. The rate limit implications:

Anthropic: Cached input tokens don’t count toward ITPM limits at all on Claude 4.x/4.5 models. With 80% cache hit rate, a 2M ITPM limit effectively handles 10M total input tokens per minute. This is a throughput multiplier, not just a cost reduction.
Groq: Cached tokens don’t count toward TPM/TPD limits on the free tier, effectively expanding your daily token budget.
Moonshot (Kimi): Automatic 75% caching discount applied without opt-in. No cache management required.
Groq: Cached tokens don’t count toward TPM/TPD limits on the free tier, effectively expanding your daily budget.

Batch APIs: higher queue limits

OpenAI, Anthropic, and Google all offer batch endpoints with queue limits that far exceed real-time TPM. OpenAI’s GPT-5.5 gets 1.5M batch queue tokens at Tier 1 vs. 500K real-time TPM. For workloads that don’t need sub-second responses, batch endpoints let you move more tokens through tighter rate limits. See batch API cost savings for pricing details.

Model routing for rate limit distribution

Routing requests to different models by complexity distributes rate limit pressure across multiple buckets instead of concentrating it on your most constrained tier. A three-tier architecture (budget for triage, mid-tier for generation, flagship for reasoning) also cuts costs 60–85%. Full routing architecture with cost modeling is in the token costs post.

Custom silicon for fewer concurrent connections

Cerebras (~2,100 tok/s) and SambaNova (~0.2s TTFT) are faster than GPU-based providers by a large margin. A task that requires 10 parallel GPU-based API calls to meet a latency SLA might need 2–3 calls on Cerebras. Fewer calls means less rate limit pressure.

Watch for shared pools and hidden constraints

Anthropic model pools: Opus 4.7 and 4.6 share one rate limit bucket. Sonnet 4.6 and 4.5 share another. Sending traffic to both model versions doesn’t double your effective limits; it draws from the same pool. Plan your model selection accordingly.
Google’s invisible limits: With static rate limit tables removed from public docs, you must check AI Studio (aistudio.google.com/rate-limit) to see your actual per-project limits. Don’t assume the numbers from third-party blog posts are current.
DeepSeek’s dynamic ceiling: No published RPM/TPM means no guaranteed minimum. Under heavy load, your effective concurrency can drop without warning. Build fallback routing to a second provider if DeepSeek availability is critical to your pipeline.