---
title: "AWS Bedrock Pricing: Every Model, Tier and Region Rule"
description: "AWS Bedrock pricing verified 2026-10-23: Claude, GPT-6 Astra, Nova and open-weight rates per 1M tokens, the 10% global vs in-region rule, Flex/Priority/Batch/Reserved tiers, caching and cost per task."
canonical: https://vynaris.com/blog/aws-bedrock-pricing
author: "Dhrumil Barot"
author_bio: "Dhrumil is an AI inference and LLM infrastructure expert. Dhrumil built Vynaris to make model access, routing and pricing simple and transparent for developers."
published: 2026-10-23
updated: 2026-10-23
---

# AWS Bedrock Pricing: Every Model, Tier and Region Rule

Published 2026-10-23 · By [Dhrumil Barot](https://vynaris.com/about#founder), Founder

AWS Bedrock pricing verified 2026-10-23: Claude, GPT-6 Astra, Nova and open-weight rates per 1M tokens, the 10% global vs in-region rule, Flex/Priority/Batch/Reserved tiers, caching and cost per task.

**TL;DR:** AWS Bedrock pricing on 2026-10-23 runs from $0.035 input / $0.14 output per million tokens (Amazon Nova Micro) to $10 / $50 (Claude Fable 5.1 and GPT-6 Astra). Claude on Bedrock with a global inference profile costs exactly the Anthropic list price; in-region and geographic endpoints add 10%. Batch and Flex halve the rate; Priority costs 1.75x.

Amazon Bedrock pricing is pay per token: each model has a USD price per 1 million input tokens and output tokens, plus cache and tier rates, billed on your AWS invoice. Prices verified 2026-10-23 for us-east-1 from the [AWS Price List API](https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrockFoundationModels/current/us-east-1/index.json) behind the [Amazon Bedrock pricing page](https://aws.amazon.com/bedrock/pricing/). On a task of 10,000 input tokens and 1,000 output tokens, one call costs $0.00049 on Nova Micro, $0.0015 on Claude Haiku 5.5, $0.03 on Claude Sonnet 5.5 and $0.06 on Claude Opus 5.5. The biggest lever most teams miss is not the model: it is the inference profile, which moves every Claude and GPT-6 Astra price by 10%.

## How much does AWS Bedrock cost per million tokens?

Bedrock prices each model per 1 million tokens on the Standard tier. These are the Claude and OpenAI models on Bedrock in us-east-1, with the global cross-Region profile price and the in-region (or geographic) price:

| Model | Global input / output | In-region input / output | Cache read (global) |
| --- | --- | --- | --- |
| Claude Fable 5.1 | $10.00 / $50.00 | $11.00 / $55.00 | $0.25 |
| GPT-6 Astra | $10.00 / $50.00 | $11.00 / $55.00 | $1.00 |
| Claude Opus 5.5 | $4.00 / $20.00 | $4.40 / $22.00 | $0.20 |
| Claude Sonnet 4.6 | $3.00 / $15.00 | $3.30 / $16.50 | $0.30 |
| Claude Sonnet 5.5 | $2.00 / $10.00 | $2.20 / $11.00 | $0.10 |
| Claude Haiku 4.5 | $1.00 / $5.00 | $1.10 / $5.50 | $0.10 |
| Claude Haiku 5.5 | $0.10 / $0.50 | $0.11 / $0.55 | $0.01 |

The global prices match Anthropic's own API rate card line for line: Opus 5.5 at $4 / $20, Sonnet 5.5 at $2 / $10, Haiku 5.5 at $0.10 / $0.50 (see our [Claude API pricing breakdown](https://vynaris.com/blog/claude-api-pricing)). GPT-6 Astra at $10 / $50 global matches [OpenAI's direct price](https://vynaris.com/blog/openai-api-pricing). So Bedrock is not a markup on these models unless you pin a region. Claude Haiku 5.5 has a long-context rate of $0.50 / $2.50 above its prompt-size threshold, and GPT-6 Astra long context costs $20 / $75 global.

## What do Amazon Nova and open-weight models cost on Bedrock?

Amazon Nova and open-weight models are Bedrock's cheapest options, priced per 1 million tokens on the Standard tier, in-region in us-east-1:

| Model | Input / 1M | Output / 1M | Cost per 10K-in / 1K-out call |
| --- | --- | --- | --- |
| Amazon Nova Micro | $0.035 | $0.14 | $0.00049 |
| Amazon Nova Lite | $0.06 | $0.24 | $0.00084 |
| gpt-oss-120b | $0.15 | $0.60 | $0.0021 |
| Qwen3 235B A22B 2507 | $0.22 | $0.88 | $0.00308 |
| Llama 4 Maverick 17B | $0.24 | $0.97 | $0.00337 |
| MiniMax M2.5 | $0.30 | $1.20 | $0.0042 |
| Nova 2 Lite | $0.33 | $2.75 | $0.00605 |
| Mistral Large 3 | $0.50 | $1.50 | $0.0065 |
| DeepSeek V3.2 | $0.62 | $1.85 | $0.00805 |
| Llama 3.3 70B | $0.72 | $0.72 | $0.0079 |
| Amazon Nova Pro | $0.80 | $3.20 | $0.0112 |
| GLM 5 | $1.00 | $3.20 | $0.0132 |
| Nova 2 Pro | $1.375 | $11.00 | $0.02475 |
| Grok 4.7 | $2.20 | $6.60 | $0.0286 |
| Amazon Nova Premier | $2.50 | $12.50 | $0.0375 |
| Kimi K3 | $3.30 | $16.50 | $0.0495 |

Nova 2 Lite, Nova 2 Pro, Grok 4.7 and Kimi K3 also have global profiles at about 10% less: Nova 2 Lite $0.30 / $2.50, Nova 2 Pro $1.25 / $10, Grok 4.7 $2 / $6, Kimi K3 $3 / $15. Llama 3.3 70B charges the same $0.72 for input and output, which makes it relatively cheap for long generations and expensive for long prompts. Look up model IDs in our [/models catalog](https://vynaris.com/models).

## What is the difference between global and in-region Bedrock pricing?

A global inference profile is a Bedrock endpoint that lets AWS route your request to any commercial Region with capacity; in-region and geographic profiles (US, EU, APAC) keep processing inside a boundary. Per the [cross-Region inference docs](https://docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html), global gives "approximately 10% savings" and there is no additional routing charge.

In the price list the premium is exactly 10% on every Claude token category: Opus 5.5 input is $4.00 global and $4.40 regional, a cache read is $0.20 and $0.22. On a workload of 1 million Opus 5.5 calls at 10,000 input and 1,000 output tokens, that is $60,000 global versus $66,000 regional, $6,000 for a model ID string. Choose geographic only when data residency or a compliance rule requires it. Our cost model for [fallback chains during outages](https://vynaris.com/blog/model-fallback-chains-outage-cost) covers the availability side of the same decision.

## What do Bedrock Flex, Priority, Batch and Reserved tiers cost?

Bedrock has four service tiers for real-time inference plus Batch; you choose one with the `service_tier` parameter set to `reserved`, `priority`, `default` or `flex` on the runtime API, per the [Bedrock service tiers docs](https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html):

| Tier | Price vs Standard | Example: Nova Pro input / output | Use for |
| --- | --- | --- | --- |
| Flex | 0.5x | $0.40 / $1.60 | evals, summarization, background agents |
| Standard (`default`) | 1x | $0.80 / $3.20 | everyday traffic |
| Priority | 1.75x | $1.40 / $5.60 | customer-facing paths that need the fastest response |
| Batch | 0.5x | $0.40 / $1.60 | async jobs on S3 input files |
| Reserved | contract | via AWS account team | 24x7 workloads that cannot tolerate throttling |

Priority, Standard and Flex share your on-demand quota; Reserved capacity is separate. Reserved requires at least 100,000 input tokens per minute and 10,000 output tokens per minute, and it bills until you delete the reservation. When sizing it, count both InputTokenCount and CacheWriteInputTokens in CloudWatch. Not every model supports every tier; Claude Sonnet 4.6 and Haiku 4.5 list Batch at exactly half price ($1.50 / $7.50 and $0.50 / $2.50 global). For the decision rule on when latency is worth paying for, see [batch APIs: when latency is free](https://vynaris.com/blog/batch-apis-openai-vs-anthropic-when-latency-is-free).

## How does Bedrock prompt caching pricing work?

Prompt caching on Bedrock bills tokens read from cache at a reduced cache-read rate, and tokens written to cache at a write rate that can exceed the input rate. For Claude the multipliers match Anthropic's: a cache read is 10% of input on Sonnet 4.6 and Haiku 4.5, and 5% on Opus 5.5 and Sonnet 5.5 ($0.20 against $4.00, $0.10 against $2.00). A 5-minute write costs 1.25x input (Opus 5.5: $5.00) and a 1-hour write costs 2x ($8.00). Amazon Nova cache reads are 75% below the input price.

You place cache checkpoints in the prompt; per the [Bedrock prompt caching docs](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html), the prefix before each checkpoint must meet a model minimum between 512 and 4,096 tokens, the TTL is 5 minutes or 1 hour, and it resets on each hit. Worked example on Opus 5.5 global: a 20,000-token system prompt costs $0.08 per uncached call. Writing it once with a 5-minute TTL costs $0.10, and each hit costs $0.004. Two calls within five minutes already beat no caching ($0.104 vs $0.16). The break-even math is in [prompt caching break-even](https://vynaris.com/blog/prompt-caching-break-even-2026).

## What does a Bedrock task actually cost?

Cost per task is what goes in a budget. Assumptions: 10,000 input tokens, 1,000 output tokens, no caching, Standard tier, us-east-1. Formula: input/1M x input price + output/1M x output price. Edit the token counts for your workload.

| Model | Global profile, per call | In-region, per call | Per 1,000 calls (global) |
| --- | --- | --- | --- |
| Claude Haiku 5.5 | $0.0015 | $0.00165 | $1.50 |
| Claude Haiku 4.5 | $0.015 | $0.0165 | $15.00 |
| Claude Sonnet 5.5 | $0.030 | $0.033 | $30.00 |
| Claude Sonnet 4.6 | $0.045 | $0.0495 | $45.00 |
| Claude Opus 5.5 | $0.060 | $0.066 | $60.00 |
| Claude Fable 5.1 | $0.150 | $0.165 | $150.00 |
| GPT-6 Astra | $0.150 | $0.165 | $150.00 |

Two results matter. Claude Haiku 4.5 costs 10x Claude Haiku 5.5 on this shape ($15 vs $1.50 per 1,000 calls), so anyone still on 4.5 for classification or extraction is overpaying. And Sonnet 4.6 costs 50% more than Sonnet 5.5. Run your own token mix through the [Vynaris cost calculator](https://vynaris.com/calculator) to compare Bedrock with direct APIs.

## What other charges appear on a Bedrock bill?

Bedrock bills features outside inference separately, and these lines often surprise teams that budgeted tokens only. us-east-1 prices from the price list:

- **Guardrails**: content filters and denied topics $0.15 per 1,000 text units each; sensitive-information filters $0.10; contextual grounding checks $0.10; word filters free. A guardrail with content and topic policies on both input and output is charged on each pass.
- **Bedrock web search**: $12 per 1,000 queries.
- **Intelligent Prompt Routing**: $1.00 per 1,000 routed requests, on top of the tokens of the model it picks.
- **Knowledge Bases**: the vector store (for example Amazon OpenSearch Serverless or Aurora) is billed by that service, plus embedding tokens and the generation model's tokens.
- **Model customization and Custom Model Import**: training per token processed and storage per month, then inference at provisioned or on-demand rates.
- **Data transfer and CloudWatch logs**: model invocation logging to S3 or CloudWatch is billed at those services' rates.

Guardrail charges scale with characters screened, not tokens, and apply to input and output separately. On a cheap model like Haiku 5.5, a guardrail on every call can cost the same order of money as the tokens themselves, so price it per call before turning it on globally.

## Is Bedrock cheaper than calling Anthropic or OpenAI directly?

Bedrock is the same price as the direct API for Claude and GPT-6 Astra when you use a global inference profile, and 10% more in-region. The reasons to choose it are not price: it bills on an existing AWS invoice and counts toward AWS commitments (EDP or private pricing), keeps traffic inside your VPC through PrivateLink, uses IAM for access control, and gives a single API for Claude, Nova, Llama, DeepSeek and other providers. The reasons to go direct: new models and features (for example beta tools or fast modes) often reach the provider's own API first, and Batch support on Bedrock varies by model. Our [LiteLLM vs OpenRouter vs managed gateways](https://vynaris.com/blog/litellm-vs-openrouter-vs-managed-gateways-tco) post covers the gateway side of this trade.

## When is the expensive Bedrock model worth it?

Claude Opus 5.5 at $60 per 1,000 calls is worth it over Sonnet 5.5 at $30 when Sonnet needs more than about 2x the tokens or attempts to finish the same task. That happens on long-horizon coding and multi-step agents. It rarely happens on extraction, classification, routing and summarization, where Claude Haiku 5.5 at $1.50 per 1,000 calls or Nova Micro at $0.49 is the right default. Measure success rate and tokens per task on your own eval set before choosing by list price; the cheapest model per token is not always the cheapest per completed task.

## Where Vynaris fits

Vynaris is an OpenAI-compatible API that routes each request to the cheapest model that can handle it across providers, including the Claude, Llama and DeepSeek models Bedrock also hosts. If your traffic mixes easy and hard requests, routing the easy share to a Haiku 5.5 or Nova-class model is where savings come from. See [Vynaris pricing](https://vynaris.com/pricing). If you are bound to an AWS commitment or need PrivateLink, stay on Bedrock and use a global profile; you don't need us for that.

## FAQ

**How is AWS Bedrock billed?**

On-demand Bedrock is billed per input and output token, per model, on your monthly AWS bill. Each response reports `inputTokens` and `outputTokens` in its usage block; multiply each by the per-1M rate and divide by 1,000,000.

**What is the cheapest model on AWS Bedrock?**

Amazon Nova Micro at $0.035 input / $0.14 output per 1M tokens in us-east-1; Batch halves it to $0.0175 / $0.07. Among Claude models, Haiku 5.5 at $0.10 / $0.50.

**How much does Claude cost on AWS Bedrock?**

With a global profile, the same as Anthropic's API: Opus 5.5 $4 / $20, Sonnet 5.5 $2 / $10, Haiku 5.5 $0.10 / $0.50 per 1M tokens. In-region and geographic endpoints cost 10% more.

**Does Bedrock have a free tier?**

Bedrock model inference has no permanent free tier; you pay per token from the first call. New AWS accounts may receive promotional credits that apply to Bedrock.

**What is Bedrock provisioned throughput?**

Reserved capacity billed per time unit instead of per token. The current form is the Reserved tier, with a minimum of 100,000 input and 10,000 output tokens per minute, arranged through your AWS account team.

**Is Bedrock Batch inference 50% cheaper?**

Yes, for supported models. Claude Sonnet 4.6 Batch is $1.50 / $7.50 against $3 / $15 on demand.

## About the author

[Dhrumil Barot](https://vynaris.com/about#founder), Founder. Dhrumil is an AI inference and LLM infrastructure expert. Dhrumil built Vynaris to make model access, routing and pricing simple and transparent for developers. [LinkedIn](https://www.linkedin.com/in/bdhrumil) · [X](https://x.com/fyidhrumil)