What is AWS Bedrock
Amazon Bedrock is AWS’s managed generative AI platform for companies that want access to multiple foundation models, governance controls, and production deployment inside the AWS operating model. It is not one model family; it is a marketplace and orchestration layer for AI work inside AWS.
That distinction matters because Bedrock is bought like infrastructure. Buyers care about provider access, guardrails, knowledge bases, model evaluation, agent capabilities, and regional deployment options rather than about one branded assistant experience.
For teams already standardised on AWS, Bedrock is often attractive because it keeps AI close to the rest of the cloud stack. It gives them access to models from providers such as Anthropic, Meta, Mistral, Amazon, and others without forcing a separate vendor relationship for every experiment.
Core offerings
- Managed access to a broad set of foundation models from multiple providers.
- Inference through Standard, Flex, Priority, and Reserved service tiers.
- Knowledge Bases, Guardrails, Agents, Prompt Optimization, Intelligent Prompt Routing, and model evaluation tooling.
- Batch inference and selected discounted workload paths for cost-sensitive jobs.
- AWS-native deployment posture for teams that want AI inside their existing cloud security and operations model.
Pricing
As of August 2, 2026, Amazon Bedrock’s public pricing page emphasises that pricing varies by model provider, modality, and service tier rather than by one flat platform subscription. AWS currently lists Standard, Flex, Priority, and Reserved service tiers for supported models. Standard uses regular model rates, Flex offers discounted standard pricing for non-urgent workloads, Priority adds a premium for faster latency, and Reserved is priced as fixed committed capacity.
AWS also publicly says select foundation models support batch inference at 50% lower price than on-demand inference. That is commercially important because Bedrock can be much more attractive for summarisation, evaluation, and asynchronous workflows than it first appears if your workloads can tolerate delay.
Bedrock’s pricing page gives model-level examples rather than one platform fee. For example, AWS’s current public page shows Anthropic Claude Sonnet 5 promotional launch pricing at US$2 per million input tokens and US$10 per million output tokens through August 31, 2026, after which AWS says standard pricing becomes US$3 and US$15. The page also lists provider-specific rates for model access, caching, and batch usage across supported vendors.
Current commercial structure
The most useful way to buy Bedrock is by understanding its service tiers and workload economics, not by looking for one flat platform sticker price.
| Layer | Public pricing signal | What it is for | Buyer note |
|---|---|---|---|
| Standard tier | Regular published model rates | Everyday production inference | Default path across all Bedrock foundation models |
| Flex tier | Discounted standard model rate | Non-urgent jobs and cost-sensitive workloads | Good fit for evaluations and multi-step workflows |
| Priority tier | Standard rate plus premium | Latency-sensitive or customer-facing workloads | AWS says many supported models can see up to 25% better OTPS latency |
| Reserved tier | Committed capacity, billed monthly | Predictable throughput and guaranteed tokens-per-minute capacity | Relevant for steady mission-critical traffic |
| Batch inference | 50% lower than on-demand on select models | Summaries, evaluations, and delayed workloads | Often one of Bedrock’s most attractive economic levers |
Why select AWS Bedrock
Bedrock is strongest for companies that already think in AWS terms and want AI inside the same governance, networking, billing, and deployment model as the rest of their cloud estate. It is especially compelling when you want provider choice without negotiating separate platform relationships for every model family.
The trade-off is that Bedrock is not the simplest product to explain to non-technical stakeholders. It behaves more like a platform and commercial framework than a single AI app. But for infrastructure-minded teams, that is precisely the appeal.