Rate Limits Errors Across All AI Providers (2026) — Complete Directory — AI Error Hub

Category hub · Cross-provider · Rate limits

Rate Limits Errors Across All AI Providers — 2026 Guide

Rate limits are the #1 error class in production AI. This is the cross-provider directory: every 429 pattern we've documented across every provider, IDE tool, and gateway — grouped so you can find your specific error in seconds.

Curated by Zain Khan

Editor, AI Error Hub

Updated September 23, 2026

Rate limits are the single most common error class in production AI applications. Every provider structures them differently — Anthropic splits input and output token metering; OpenAI combines them into TPM; Google uses per-project quotas; Bedrock is per-region with cross-region profiles; Azure has PTUs. This page is the cross-provider directory: every rate-limit error we've documented across every provider, IDE tool, and gateway, grouped so you can find your specific error in seconds.

For a deep-dive on rate limit architecture, backoff, and circuit breakers, see the pillar guide on handling rate limits in AI APIs. For cost optimization patterns that reduce rate-limit pressure, see the Cost Optimization Playbook.

Anthropic Claude OpenAI Google Gemini AWS Bedrock Azure OpenAI IDE Tools Gateway Layer Framework Layer

2026 rate limits at a glance

Rate limits vary by provider and tier. Use this table as a quick reference before diving into specific error patterns.

Provider	Tier	RPM	TPM/ITPM	Notes
Anthropic Claude	Tier 1	50 RPM	50K ITPM / 10K OTPM	Auto-progression after $5 payment
Anthropic Claude	Tier 4	4,000 RPM	400K ITPM / 80K OTPM	$400+ payment history
OpenAI	Tier 1	500 RPM	30K TPM (GPT-5.6)	Auto after $5 payment + 7 days
OpenAI	Tier 5	10,000 RPM	2M TPM	$1,000+ paid, 30+ days
Google Gemini (Vertex)	Paid tier	300 RPM	4M TPM	Per-project quota; can request increase
AWS Bedrock	On-demand	Model-specific	Model-specific	Cross-region profiles multiply capacity
AWS Bedrock	Provisioned Throughput	Reserved capacity	Reserved capacity	Fixed cost, guaranteed throughput
Azure OpenAI	Standard	Per-deployment	Per-deployment	Per-deployment TPM allocation
Azure OpenAI	PTU	Guaranteed	Guaranteed	100 PTU minimum for GPT-5.6 Terra Global

Anthropic Claude

Anthropic pioneered the input/output token split. Claude enforces four meters simultaneously per API key: RPM, ITPM (input tokens/min), OTPM (output tokens/min), and TPD on lower tiers. Auto-progression from Tier 1 to Tier 4 based on payment history. Batches API has its own quota.

→ Claude API Pillar Guide

OpenAI

OpenAI uses TPM (input + output combined) and RPM per model. Each model has independent limits — GPT-5.6 Terra has separate limits from Sol, from Luna. Batch API has separate quotas. The Responses API uses the same underlying limits as chat completions.

→ OpenAI API Pillar Guide

Google Gemini

Gemini uses per-project 'quotas' rather than tier-based rate limits. Separate quota for generate_content, embed_content, and file uploads. Free tier (AI Studio) has aggressive limits; paid tier (Vertex AI) has higher limits with a straightforward request path.

→ Gemini API Pillar Guide

AWS Bedrock

Bedrock quotas are per-model, per-region. Quota increases via AWS Service Quotas requests. The 2026 landscape added cross-region inference profiles that aggregate US/EU/APAC regions transparently for higher throughput without PT commitment.

→ Amazon Bedrock Pillar Guide

Azure OpenAI

Azure has two paths: Standard (pay-as-you-go) with TPM/RPM per deployment, and Provisioned (PTUs) for guaranteed throughput. PTU minimums are model-specific — GPT-5.6 Terra requires 100 PTUs minimum for Global deployment.

→ Azure OpenAI Pillar Guide

Rate limits at other layers of the stack

Rate limits don't only fire at the provider layer. IDE tools, gateways, and agent frameworks all add their own throttling — often for different reasons and with different failure modes.

IDE Tools

IDE coding tools (Cursor, Claude Code, GitHub Copilot) add their own rate limit layer on top of the underlying provider limits. When you hit one, understanding which layer fired is critical — upgrading the direct API key does not help if the tool subscription tier is the bottleneck.

→ IDE Tool Comparison Post

Gateway Layer

Gateways (LiteLLM, OpenRouter, Portkey) add another rate limit layer. Their rate limits are independent from the underlying provider's — sometimes higher, sometimes lower, sometimes structured completely differently.

→ AI Gateway Comparison Post

Framework Layer

Agent frameworks (LangChain, LangGraph, LlamaIndex) don't have their own rate limits, but they generate more requests per user turn than direct API calls do. Rate limit management in frameworks is about controlling iteration count and caching intermediate results.

→ LangChain Pillar Guide

Frequently asked questions

What causes AI API rate limits?

AI APIs meter multiple resources simultaneously: requests per minute (RPM), tokens per minute (TPM), and increasingly input tokens per minute (ITPM) and output tokens per minute (OTPM) as separate quotas. Each meter protects a different scarce resource — GPU throughput, request queue length, sustained daily load. Any one can trigger a 429.

How do I fix a 429 rate limit error?

Three steps: (1) check the response body to identify which meter fired — RPM, TPM, or ITPM/OTPM. (2) Implement exponential backoff with full jitter and respect the retry-after header. (3) Address root cause — either upgrade tier, reduce token count via prompt caching, or reduce request rate via client-side throttling.

Which AI provider has the best rate limits?

It depends on your workload. OpenAI's top tier reaches 10,000 RPM / 2M TPM. Claude's Tier 4 hits 4,000 RPM / 400K ITPM. Gemini's paid tier offers 300 RPM / 4M TPM (large TPM, moderate RPM). For guaranteed capacity, Bedrock Provisioned Throughput and Azure PTUs are dedicated reserved allocations.

How much does provisioned throughput cost?

Bedrock PT and Azure PTUs are priced per unit per hour, with monthly or annual commitments for discounted rates. Break-even against on-demand is typically 60-70% sustained utilization. Below that, on-demand plus horizontal scaling is cheaper. See our Cost Optimization Playbook for the full math.

Can I share rate limits across multiple API keys?

It varies by provider. Most providers meter rate limits per-key OR per-org (whichever is lower). Splitting traffic across N keys within the same org often doesn't multiply capacity if there's an org-level cap. Keys on separate orgs do give independent capacity. Check your provider's TOS before scaling multi-key setups.

Why do I get rate-limited on my first request of the day?

Two possible causes. (1) Rate limit windows are typically sliding, not fixed — yesterday's traffic still counts if within the window. (2) Tool-specific rate limits (e.g., GitHub Copilot Agent Mode session limits) are time-based, not per-chat. Opening a new session doesn't reset them.

How do I request a rate limit increase?

Every major provider supports increase requests: Anthropic via the Console, OpenAI via usage tier progression or direct request, Google Cloud via IAM & admin quota console, AWS via Service Quotas, Azure via PTU purchase or standard deployment increase. Typical turnaround is 1-5 business days.

What's the difference between rate limits and quotas?

Terminology varies by provider. 'Rate limit' typically means per-minute enforcement (RPM/TPM). 'Quota' typically means longer-window enforcement (per-day, per-month, or lifetime). Google uses 'quota' as an umbrella term for both. Functionally, both are enforced the same way (429 response with retry-after).