What it is
Amazon Bedrock is AWS's service for calling foundation models without running servers. One API reaches models from Anthropic, Meta, Mistral, Amazon, OpenAI (gpt-oss), DeepSeek, Qwen, Moonshot (Kimi) and others, and Bedrock adds fine-tuning, agents, knowledge bases and guardrails around them. You can also import weights you trained elsewhere with Custom Model Import.
Most use is billed per token. According to the pricing page, Qwen3 Next 80B costs $0.15 input and $1.20 output per million tokens in US regions, and DeepSeek V3.2 $0.62 and $1.85. Batch jobs and the Flex tier cost 50% less on selected models, a Priority tier costs 75% more, and provisioned throughput reserves capacity by the hour. Where a model offers both, US-only cross-region inference costs about 10% more than global routing: Kimi 3 is $3.00 and $15.00 globally against $3.30 and $16.50 in the US.
Anthropic's newest Claude Mythos models (Mythos 5.1, Mythos 5 and Mythos Preview) are gated on Bedrock and need approval through an Anthropic account team.
Our review
Bedrock's value is less the models than where they run. According to the FAQ, inputs and outputs are not used to train models by AWS or the model providers and are not shared with those providers, data is encrypted in transit and at rest with optional customer-managed keys, and PrivateLink keeps traffic off the public internet. The compliance list is long: SOC 1, 2 and 3, ISO 27001, 27017, 27018 and 27701, FedRAMP Moderate, HIPAA eligibility and CSA STAR Level 2. For a regulated company already buying from AWS, that list is most of a vendor security review done.
The quota system catches teams out more often than the price. Bedrock throttles by tokens per minute, and the docs say output tokens from Claude Opus 5.5, Sonnet 5 and Fable 5.1 count ten times against that quota (Claude 4.8 counts fifteen times), while you are billed only for the real tokens. A request also reserves its full max_tokens up front, so a generous max_tokens on a busy endpoint throttles long before the bill looks large. AWS warns that new or low-history accounts can get quotas below the published defaults, and increases are not automatic.
Open-model prices sit close to specialist hosts. DeepSeek V3.2 at $0.62 and $1.85 per million tokens and Qwen3 Next 80B at $0.15 and $1.20 are competitive, and Batch and Flex halve them for work that can wait. Claude list prices did not render on the pricing page when we checked, and provisioned throughput for Anthropic models goes through an account team. Teams budgeting for Claude should take prices from the AWS console or a quote rather than assume parity with Anthropic's own API.
Bringing your own model is narrower than on a GPU platform. Custom Model Import charges $0.05718 per model-unit minute in US East and West, billed in five-minute windows plus $1.95 per unit per month for storage, and the FAQ lists only Llama, Mistral and Flan architectures (OpenAI-format imports cost $0.1433 a minute). Modal, Baseten and Runpod accept arbitrary serving stacks; Bedrock accepts a short list in exchange for never touching a GPU.
The lock-in sits in the extras rather than the models. Guardrails at $0.15 per 1,000 text units for content filters, Knowledge Bases at $5 per GB a month plus $1 per 1,000 retrievals, and Bedrock Agents are AWS-specific APIs. The Converse API is easy to swap out later; a retrieval pipeline built on Knowledge Bases is not.
Where it earns its place
A platform team at a US health insurer already on AWS
Put Claude behind claims-summary tooling without a new vendor review
Calls go through Bedrock in the insurer's own account over PrivateLink, within Bedrock's HIPAA eligibility under the insurer's AWS BAA. The team sets max_tokens tightly and requests a quota increase early, because Claude output counts 10x against tokens-per-minute quotas.
A data engineer reprocessing two million support tickets a month
Classify and tag tickets overnight at the lowest cost
A Batch job on Qwen3 Next 80B runs at half the on-demand $0.15 input and $1.20 output per million tokens. Results land in S3 by morning; nothing needs to answer in real time, so the 50% discount costs nothing in practice.
An ML engineer with a fine-tuned Llama 3 checkpoint
Serve the custom model without running GPUs
Custom Model Import loads the weights and bills $0.05718 per model-unit minute in US East, scaling to zero when idle, plus $1.95 a month storage. A Qwen or DeepSeek fine-tune would not import; those need a GPU host.
A procurement lead at a FedRAMP-bound software vendor
Approve a generative AI service for government customers
Bedrock is in AWS's FedRAMP Moderate, SOC and ISO 27001 scope, and prompts are not shared with model providers. The remaining check is per model: availability and price vary by region, and Mythos models need separate Anthropic approval.
Before you commit
- Setup
- Enable model access in the Bedrock console, grant IAM permissions and call the Converse API through any AWS SDK; first response in minutes for an existing AWS account.
- Integrations
- AWS SDKs, OpenAI-compatible Chat Completions and Responses APIs, PrivateLink, CloudWatch, S3, Lambda; Guardrails, Knowledge Bases and Agents as managed add-ons.
- Data and privacy
- No training on inputs or outputs by AWS or providers; encryption at rest with optional KMS keys; invocation logging off by default and written only to your S3 or CloudWatch; SOC 1/2/3, ISO 27001, FedRAMP Moderate, HIPAA eligible.
- Cost at scale
- Production cost is tokens plus extras: Guardrails content filters add $0.15 per 1,000 text units, Knowledge Bases $5 per GB a month and $1 per 1,000 retrievals; Batch or Flex halves token cost for deferrable work, and Priority adds 75%.
- Limits
- Token-per-minute quotas per model and endpoint; Claude output burns 10x to 15x, max_tokens is reserved at request start, and quotas for new accounts can sit below defaults until an increase is approved.
What's good
- Claude, Llama, DeepSeek, Qwen, Kimi and Nova under one AWS bill and IAM
- Inputs and outputs not used for training or shared with model providers
- SOC 1/2/3, ISO 27001, FedRAMP Moderate, HIPAA eligible; PrivateLink
- Batch and Flex tiers at half price on supported models
- Invocation logging is off by default and writes only to your own S3 or CloudWatch
What's not
- Claude output tokens burn 10x to 15x against token quotas
- New accounts may get quotas below the published defaults
- Custom Model Import supports only Llama, Mistral and Flan architectures
- Claude provisioned throughput and Mythos access need an account team
- Guardrails, Knowledge Bases and Agents tie applications to AWS APIs
Best for
Companies already on AWS that need Claude or open models with enterprise data controls and one bill.
Alternatives
Amazon Bedrock: common questions
- Is Amazon Bedrock worth it?
- For a company whose data already lives in AWS, Bedrock turns Claude and the major open models into another IAM-governed service; for everyone else it is a long pricing page and a quota system to learn. We rate it 4.3 out of 5.
- How much does Amazon Bedrock cost?
- Per 1M tokens on demand, e.g. DeepSeek V3.2 $0.62 in / $1.85 out (US), Qwen3 Next 80B $0.15 / $1.20 (US). Batch and Flex 50% off; Priority +75%. Provisioned throughput per model unit per hour. Custom Model Import $0.05718 per model-unit minute (US East and West). Guardrails from $0.10 per 1,000 text units.
- What are the downsides of Amazon Bedrock?
- Claude output tokens burn 10x to 15x against token quotas; new accounts may get quotas below the published defaults; custom Model Import supports only Llama, Mistral and Flan architectures; claude provisioned throughput and Mythos access need an account team; guardrails, Knowledge Bases and Agents tie applications to AWS APIs.
- Who is Amazon Bedrock best for?
- Companies already on AWS that need Claude or open models with enterprise data controls and one bill.
- How much does Amazon Bedrock cost for a team or in production?
- Production cost is tokens plus extras: Guardrails content filters add $0.15 per 1,000 text units, Knowledge Bases $5 per GB a month and $1 per 1,000 retrievals; Batch or Flex halves token cost for deferrable work, and Priority adds 75%.
- Is Amazon Bedrock safe for company data?
- No training on inputs or outputs by AWS or providers; encryption at rest with optional KMS keys; invocation logging off by default and written only to your S3 or CloudWatch; SOC 1/2/3, ISO 27001, FedRAMP Moderate, HIPAA eligible.
- What are the best Amazon Bedrock alternatives?
- The closest alternatives we have reviewed are Google Agent Platform (Vertex AI), Together AI, Baseten and OpenRouter.
Answers checked Oct 22, 2026.
Make Amazon Bedrock? Show this rating on your site
