Amazon Bedrock Claude rate limits: quotas, throttling and increases

The quotas that apply to Claude

Despite the names, these quotas cover every inference API that calls the model on bedrock-runtime, including Converse, Responses and Chat Completions. The newer bedrock-mantle endpoint is counted separately, with separate input and output tokens-per-minute quotas and no RPM quota.

Published defaults for current Claude models

Why the quota drains faster than the bill

Bedrock converts output tokens into quota usage at a burndown rate. AWS lists 10x for Claude Opus 5.5, Sonnet 5, Opus 5 and Fable 5.1, 15x for Claude version 4.8 models, and 5x for Claude models version 4.7 and below. The page names no separate rate for Sonnet 5.5 or Haiku 5.5. You are billed only for the tokens actually used.

The quota is also charged up front. When a request starts, Bedrock deducts the input tokens plus the full max_tokens value, and returns the unused part when the request ends. Cache reads don't count toward the quota. Applying AWS's formula to one Claude Opus 5.5 request:

input tokens        20,000
max_tokens          32,000
output generated     2,000

reserved at start   52,000
  = input + max_tokens
charged at the end  40,000
  = input + output x 10
billed tokens       22,000
  = input + output

Setting max_tokens close to the real output size is the cheapest way to fit more concurrent requests under the same quota.

Cross-Region inference

Cross-Region inference sends a request through an inference profile, and Bedrock picks the Region that serves it. Calls through a profile use the cross-Region quota for the model, which is separate from the on-demand quota for single-Region calls.

ThrottlingException and the errors next to it

AWS's example retry budget is six attempts in total, the first request plus five retries. In boto3:

import boto3
from botocore.config import Config

config = Config(retries={
    "total_max_attempts": 6,
    "mode": "standard",
})
client = boto3.client(
    "bedrock-runtime", config=config
)

How to request a quota increase

1. Open Service Quotas, choose Amazon Bedrock and check the values applied to your account for the model and Region you call.
2. Check the model's lifecycle status first. AWS does not grant increases for Legacy or Deprecated models.
3. Request the increase on "Cross-Region InvokeModel tokens per minute" for the model. AWS Support then offers to raise the on-demand tokens per minute and tokens per day quotas in the same case.
4. Show real usage. AWS says priority goes to customers whose traffic already consumes their existing quota, and requests from others may be denied. Smaller increases are often approved automatically; larger ones go to AWS Support for review.
5. For the bedrock-mantle endpoint, follow its separate increase process.

# Applied Claude quotas, current Region
aws service-quotas list-service-quotas \
  --service-code bedrock \
  --query "Quotas[?contains(QuotaName,
    'Claude Sonnet 5.5')].[QuotaName,Value]" \
  --output table

# Request an increase with the quota code
aws service-quotas \
  request-service-quota-increase \
  --service-code bedrock \
  --quota-code L-XXXXXXXX \
  --desired-value 12000000

Provisioned Throughput and service tiers

Getting more from the quota you have

• Lower max_tokens to the output you expect, since the full value is reserved at the start of each request.
• Use prompt caching for repeated context.
• Limit by tokens and concurrency. AWS warns that a requests-per-minute limiter alone doesn't protect you when request sizes change.
• Ramp up in steps, and hold each level long enough to watch 429, 503 and 529 rates.
• Move offline work to batch inference, and latency-tolerant work to the Flex tier where the model supports it.

Routing past a single account's quota

AWS's own advice for meeting an availability target is to add bounded retries and multi-Region or multi-model fallback in your application. A gateway does that work for you, across several providers.

With Standard Compute, your Bedrock account becomes one of several paths for Claude traffic. Claude requests are spread across Anthropic and the major clouds, and the same API reaches GPT, Gemini and open models such as GLM, DeepSeek, MiniMax and Kimi. The router picks the cheapest model that can handle each step and the cheapest provider fast enough to serve it, by latency and tokens per second.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.stdcmpt.com/v1",
    api_key=STANDARD_COMPUTE_KEY,
)

# "standardcompute" routes each step to the
# cheapest model that can do it. To pin Claude:
# model="anthropic/claude-sonnet-5.5"
client.chat.completions.create(
    model="standardcompute",
    messages=messages,
)

For Claude Desktop and Claude Code seat limits, see Claude rate limits for teams. For a general fix list, see AI agent keeps getting rate limited.

Frequently asked questions

Sources (checked 2026-11-07)

1. Quotas for Amazon Bedrockdocs.aws.amazon.com
2. Quotas for the bedrock-runtime endpointdocs.aws.amazon.com
3. How tokens are counted in Amazon Bedrockdocs.aws.amazon.com
4. Amazon Bedrock service quotas (AWS General Reference)docs.aws.amazon.com
5. Cross-Region inferencedocs.aws.amazon.com
6. Troubleshooting Amazon Bedrock API error codesdocs.aws.amazon.com
7. Scaling and throughput best practicesdocs.aws.amazon.com
8. Provisioned Throughputdocs.aws.amazon.com
9. Service tiers for inferencedocs.aws.amazon.com
10. Requesting a quota increase (Service Quotas)docs.aws.amazon.com
11. Claude Code on Amazon Bedrockcode.claude.com

Related

• Standard Compute for enterprise
• Claude rate limits for teams
• Anthropic API rate limits by tier
• LLM API rate limits compared