Bedrock Quotas for Claude in the EU (2026)

TL;DR

Bedrock quotas for Claude are tokens per minute per model and per Region. In EU regions the default is 30 million TPM for Opus 5.5 and 6 million for Sonnet 5, but each output token burns 10 quota tokens. Classic Provisioned Throughput does not exist for current Claude models. Reserved capacity covers only Haiku 4.5.

How do Bedrock quotas for Claude work?

Bedrock quotas for Claude are token budgets per minute, set per model and per AWS Region, plus one daily ceiling per account. The runtime endpoint counts input and output against a single tokens-per-minute (TPM) quota. Some models also have a requests-per-minute (RPM) quota. Throttling starts when either is exhausted, and the defaults vary by account.

The runtime quota page names four quota types for bedrock-runtime:

Quota	Scope	What it limits
Cross-Region InvokeModel tokens per minute	Per model, per Region	Input plus output tokens through a geo or global inference profile
On-demand InvokeModel tokens per minute	Per model, per Region	Input plus output tokens for single-Region calls
InvokeModel requests per minute	Per model, per Region	Requests; enforced only for some models
Cross-Model Max Tokens Per Day	Per account, per Region	All supported models together

Three details matter for EU teams. First, the quota is shared across InvokeModel, Converse, Responses and Chat Completions. Switching API does not give you fresh capacity. Second, the bedrock-mantle endpoint has its own, separate quotas with input and output counted apart, so plan the two endpoints independently. Third, AWS says defaults can change with “regional factors, payment history, fraudulent usage”, and the mantle page warns that new accounts may receive reduced quotas. Your Service Quotas console is the truth, not a blog table.

Nearly all EU Claude traffic runs through the EU geo profile, so the cross-Region TPM quota is the number that throttles you. Which Claude models have an EU profile, and how to restrict calls to it, is in our Claude on AWS Bedrock EU guide.

What are the default quotas for Claude in EU regions?

AWS publishes defaults in the General Reference quota table as “each supported Region”, so the same values apply when you call from Frankfurt, Ireland or Paris. The Opus models start at 30 million TPM, Sonnet 5 at 6 million, Haiku 4.5 at 5 million. All TPM quotas below are adjustable.

Model	EU geo profile	Cross-Region TPM	Global TPM	Output burndown	Service tiers
Claude Opus 5.5	eu.anthropic.claude-opus-5-5	30,000,000	30,000,000	10x	Standard
Claude Opus 5	eu.anthropic.claude-opus-5	30,000,000	30,000,000	10x	Standard, Batch
Claude Opus 4.8	eu.anthropic.claude-opus-4-8	30,000,000	30,000,000	15x	Standard
Claude Sonnet 5.5	eu.anthropic.claude-sonnet-5-5	not yet published	not yet published	not yet published	Standard
Claude Sonnet 5	eu.anthropic.claude-sonnet-5	6,000,000	6,000,000	10x	Standard
Claude Haiku 4.5	eu.anthropic.claude-haiku-4-5-20251001-v1:0	5,000,000 (10,000 RPM)	5,000,000 (10,000 RPM)	5x	Standard, Reserved

Sources: quota values from the General Reference, burndown from the token counting page, tiers from each model card, all read on 1 October 2026. Sonnet 5.5 launched on Bedrock on 28 September 2026 and had no row in either table yet. Check your console before you size a launch on it.

Two more numbers belong in every capacity plan. The Cross-Model Max Tokens Per Day quota is 700 million per Region and is marked not adjustable in Service Quotas. AWS support may still offer to raise it when you request a cross-Region TPM increase. And Opus 4.7 and Opus 4.8 have no RPM quota at all. Only tokens throttle them.

The 15x factor for Opus 4.8 is not a typo in our table. AWS states 15x for “Anthropic Claude models version 4.8” and 10x for Opus 5.5, Sonnet 5, Opus 5 and Fable 5.1. Every output token from Opus 4.8 therefore drains 50% more quota than the same token from Opus 5.5. On output-heavy agents that alone is a migration argument.

How does quota consumption work? A worked example

Bedrock deducts input tokens plus max_tokens when a request starts. At the end it recalculates with real output multiplied by the burndown rate and refunds the difference. Cache reads never count. So two levers decide how many requests fit into a minute: an honest max_tokens and prompt caching.

The token counting page gives the formulas:

At request start:  input tokens + max_tokens
At request end:    InputTokenCount + CacheWriteInputTokenCount + (OutputTokenCount x burndown)
Not counted:       CacheReadInputTokenCount

Worked example, our numbers, Claude Sonnet 5 on the EU geo profile with the 6 million TPM default:

1. A support assistant sends 20,000 input tokens, no cache, max_tokens 16,000, and receives 2,000 output tokens.
2. At start Bedrock reserves 20,000 + 16,000 = 36,000 quota tokens.
3. At the end the request costs 20,000 + 2,000 x 10 = 40,000 quota tokens.
4. The bill counts 22,000 tokens: 20,000 x 2.20 USD plus 2,000 x 11.00 USD per million, so 0.066 USD at EU geo prices.
5. Capacity: 6,000,000 / 40,000 = 150 such requests per minute.

Now put 15,000 of those input tokens behind a cache hit. The final charge against quota drops to 5,000 + 2,000 x 10 = 25,000 tokens, and the same quota carries 240 requests per minute. Prompt caching is a capacity tool, not only a cost tool.

The max_tokens trap is the opposite case. Set it to 64,000 “to be safe” and each request reserves 84,000 tokens at start. A burst of parallel calls then hits the ceiling well before real usage justifies it. AWS’s own example shows the same pattern: 36,000 reserved against 9,000 actually used. Measure OutputTokenCount in CloudWatch and set max_tokens slightly above your real p99.

How do you check and raise Bedrock quotas with the AWS CLI?

You read and raise Bedrock quotas through the Service Quotas API, per Region. Increases for any Claude model go through one quota: Cross-Region tokens per minute for that model. AWS support then offers to raise on-demand TPM and the daily ceiling with it. Requests from accounts that already use their allocation get priority.

The commands below run against Frankfurt. The Service Quotas docs describe the flow as list services, list quotas, request with the quota code. Quota names in the console spell the model as “Anthropic Claude Sonnet 5”, so filter on the model name rather than hard-coding quota codes.

# 1. Find the Service Quotas code for Amazon Bedrock
aws service-quotas list-services --region eu-central-1 --query "Services[?contains(ServiceName, 'Bedrock')]" --output table

# 2. Show applied Claude Sonnet 5 quotas (replace SERVICE_CODE with the value from step 1)
aws service-quotas list-service-quotas --service-code SERVICE_CODE --region eu-central-1 --query "Quotas[?contains(QuotaName, 'Claude Sonnet 5')].[QuotaName,QuotaCode,Value,Adjustable]" --output table

# 3. Request 12M TPM on the cross-region quota (QuotaCode from step 2)
aws service-quotas request-service-quota-increase --service-code SERVICE_CODE --quota-code L-XXXXXXXX --desired-value 12000000 --region eu-central-1

# 4. Track the request
aws service-quotas list-requested-service-quota-change-history-by-quota --service-code SERVICE_CODE --quota-code L-XXXXXXXX --region eu-central-1

The CLI needs ServiceQuotasFullAccess or iam:CreateServiceLinkedRole, because the request may open a support case on your behalf. The Service Quotas guide says smaller increases are usually approved automatically, larger ones go to Support, can be partially approved and “don’t receive priority support”.

Three rules from the runtime quota page change how you plan:

1. Show traffic first. AWS states that, due to demand, priority goes to customers whose traffic consumes their existing allocation. A request from an idle account may be denied.
2. Do not ask for a retiring model. No increases for models in Legacy or Deprecated status. When we checked on 1 October 2026, Haiku 4.5 was Active, with “EOL no sooner than” 16 October 2026 on its card. Do not build a launch plan on its quota without checking the lifecycle page first.
3. Ask per Region. Quotas are scoped per Region. Raise the quota in the source Region your application actually calls.

When does Provisioned Throughput or Reserved pay off for Claude?

For current Claude models on Bedrock, classic Provisioned Throughput is not an option. The supported-models list contains only Claude 2, Claude Instant, Claude 3 and Claude 3.5. AWS adds that inference profiles do not support Provisioned Throughput. The one reservation product for a current Claude model in the EU is the Reserved tier, and only for Haiku 4.5.

The Provisioned Throughput model list shows Claude 2.1 in Frankfurt and Claude 3 Haiku and Claude 3 Sonnet in Ireland and Paris as the EU entries. Default model-unit quotas for those are 0, and AWS sends you to your account team for prices. Unless you run an old fine-tuned model, ignore it.

The service tiers page defines Reserved as capacity booked in input and output TPM for 1 or 3 months, minimum 100,000 input TPM and 10,000 output TPM, with a 99.5% uptime target and automatic overflow to Standard. It sits outside your on-demand quota. Priority and Flex share the on-demand quota, and the model cards list neither for any current Claude model.

The price arithmetic is the interesting part. Haiku 4.5 Reserved on the EU geo profile costs 0.066 USD per hour per 1,000 input TPM on a 1-month term. One hour at 1,000 TPM is 60,000 tokens, which works out to 1.10 USD per million, exactly the on-demand EU geo price. The 3-month term at 0.0594 USD is 10% cheaper. The minimum reservation costs 9.90 USD per hour, about 7,200 USD per month (our calculation). Reserved is a capacity guarantee you pay for at full utilisation, not a discount.

Situation	Our recommendation	When the other option wins
Sonnet or Opus, steady load under the default TPM	Standard on demand, EU geo profile	Never worth a reservation
Launch that will exceed default TPM	Quota increase 3 to 4 weeks ahead, with traffic history	Reserved only if the model is Haiku 4.5
Haiku 4.5, 24/7 load near or above 100,000 input TPM	Reserved 3 months if throttling hurts revenue	Standard if load drops at night or weekends
Opus 5 jobs that can wait	Batch at 50% of on-demand price	Standard when results are needed in seconds
No personal data, cost first	Global profile, about 10% cheaper per AWS	EU geo whenever GDPR residency applies
Fine-tuned Claude 3 model	Provisioned Throughput via account team	Migrate to a current model on demand

The global profile has its own TPM quota with the same default values, but it routes worldwide. For personal data under GDPR we stay on EU geo and accept the 10% premium. The full cost picture is in our Claude API pricing EU comparison.

Run it yourself or outsource?

Running Bedrock quotas in-house is a small but recurring platform job: watch token metrics, keep max_tokens honest, file increases before launches, track model lifecycles. In our estimate that is a few hours per month for a platform or FinOps engineer, plus a spike before each launch. Outsourcing pays off only when you lack AWS account ownership skills altogether.

The recurring tasks are concrete:

1. Weekly look at InputTokenCount, CacheWriteInputTokenCount and OutputTokenCount per model in CloudWatch, multiplied by the burndown rate.
2. A quota request three to four weeks before any launch or migration (our working assumption, AWS gives no SLA), with real traffic attached.
3. A lifecycle check each quarter: AWS refuses increases on Legacy models, so a migration to a newer model also resets your capacity planning.
4. For Haiku 4.5 at high load, a Reserved negotiation with the AWS account team.

An external managed service provider makes sense if nobody in-house owns the AWS organisation, or if you run several production Claude workloads and want 24/7 throttling response. It does not make sense for a single internal assistant that sits far below 6 million TPM.

If you outsource, demand these points in writing. A response time for throttling incidents and for quota requests filed on your behalf. A GDPR data processing agreement (DPA) and a current subprocessor list. An access model where the provider works through a role in your AWS account, never an account they own. Quota history and Reserved contracts belong to the account, so an exit clause must leave both with you. How the EU profile is enforced with SCPs is covered in our Amazon Bedrock EU guide.

FAQ

What are the default Bedrock quotas for Claude Sonnet 5?

6,000,000 tokens per minute for the cross-Region quota and 6,000,000 for the global quota, per Region. Each output token counts 10 times. Your account value can differ from this default, so check Service Quotas in your source Region.

Why do output tokens count 10 times against my quota?

AWS applies a burndown rate. For Opus 5.5, Sonnet 5, Opus 5 and Fable 5.1 one output token consumes 10 quota tokens, for Claude 4.8 models 15, for older Anthropic models 5. You are billed only for actual tokens.

How much does Reserved capacity for Claude cost in the EU?

Only Haiku 4.5 offers Reserved among current Claude models. On the EU geo profile it costs 0.066 USD per hour per 1,000 input TPM and 0.33 USD per 1,000 output TPM on a 1-month term. The minimum reservation is about 9.90 USD per hour.

Provisioned Throughput vs quota increase: which one do I need for Claude?

A quota increase. Provisioned Throughput by model units is listed only for Claude 2, Instant, Claude 3 and Claude 3.5, and inference profiles do not support it. For Opus 5.5, Sonnet 5.5 or Sonnet 5 you raise the cross-Region TPM quota.

Do prompt cache reads count against Bedrock quotas?

No. CacheReadInputTokenCount is excluded from the quota calculation. Cache writes count like normal input. Heavy caching therefore raises the number of requests a fixed TPM quota can carry, in our Sonnet 5 example from 150 to 240 per minute.

Does the global profile give me more quota than the EU profile?

It has its own quota with the same default values. It does not add EU-resident capacity, because global routing can process data in any commercial AWS Region. For personal data we keep the EU geo profile.

Sources

1. AWS: Quotas for Amazon Bedrock (1 October 2026)
2. AWS: How tokens are counted in Amazon Bedrock (1 October 2026)
3. AWS: Quotas for the bedrock-runtime endpoint (1 October 2026)
4. AWS: Quotas for the bedrock-mantle endpoint (1 October 2026)
5. AWS General Reference: Amazon Bedrock endpoints and quotas (1 October 2026)
6. AWS: Service tiers for optimizing performance and cost (1 October 2026)
7. AWS: Supported Region and models for Provisioned Throughput (1 October 2026)
8. AWS: Cross-Region inference (1 October 2026)
9. AWS Service Quotas: Requesting a quota increase (1 October 2026)
10. AWS: Claude Haiku 4.5 model card (1 October 2026)
11. AWS: Claude Opus 5.5 model card (1 October 2026)
12. Amazon Bedrock pricing (1 October 2026)