AWS Bedrock has become the most model-diverse managed AI service among the three major cloud providers, now offering access to 15 different foundation models from Anthropic, OpenAI, Meta, Mistral, Cohere, and Amazon itself, all through a single API. The platform's expansion in 2026 eliminates a long-standing trade-off for enterprise customers: they no longer have to choose between keeping data inside their AWS environment and accessing cutting-edge AI models.

Bedrock is essentially a gateway that turns frontier AI model APIs into AWS line items. Data never leaves your AWS infrastructure, which is critical for regulated industries like healthcare, finance, and defense that operate under HIPAA, FedRAMP, and ITAR compliance requirements. You access models through your existing AWS account, governed by your existing security tools like IAM and CloudTrail, without negotiating separate vendor contracts or building your own inference infrastructure.

What Models Are Actually Available on Bedrock Right Now?

The platform's model lineup spans multiple tiers and use cases. Anthropic's Claude family remains the most-deployed option on Bedrock, with three versions available: Claude Opus 4.7 for frontier-level tasks at $15 per million input tokens, Claude Sonnet 4.6 as the workhorse tier at $3 per million input tokens, and Claude Haiku 4.5 for speed-critical applications at $0.80 per million input tokens.

The April 2026 addition of OpenAI's GPT-5.5 family marked a significant shift. AWS historically lagged behind Azure by months on OpenAI model access, but GPT-5.5 launched on Bedrock the same week as on Azure. The lineup includes GPT-5.5 "Spud" at the frontier tier, GPT-Rosalind for multimodal vision and audio tasks, and GPT-Codex for specialized code work.

Amazon's own Nova line, launched in late 2024 and upgraded twice in 2025, offers a cost-competitive alternative. Nova Micro, the company's text-only ultra-light model, costs just $0.035 per million input tokens, making it the cheapest text model available on any hyperscaler platform. This pricing makes Nova Micro practical for high-volume classification and routing tasks where cost per inference matters more than raw capability.

Beyond these, Bedrock also supports Meta's Llama 4 family in three sizes, Mistral's models optimized for EU data residency, Cohere's enterprise-focused RAG (retrieval-augmented generation) stack, AI21's long-context Jamba 2, Stability's image and video generation tools, and DeepSeek's V3.5 model with V4 coming soon.

How Does Bedrock's Pricing Actually Work?

Bedrock offers two billing modes designed for different workload patterns. On-demand pricing charges per token with no commitment, making it suitable for bursty workloads, development work, and anything processing under roughly 10 million tokens per day. The trade-off is that you're subject to model-level rate limits during peak hours.

Provisioned throughput requires a minimum commitment of "model units" for either one month or six months, guaranteeing fixed throughput in tokens per minute. This approach works better for predictable high-volume workloads exceeding 50 million tokens per day, latency-critical applications, and situations where guaranteed capacity during business hours is essential. The math is compelling for stable workloads: a six-month provisioned commitment costs roughly 30 percent less per token than on-demand pricing, while a one-month commitment saves about 15 percent.

To illustrate the savings, consider a workload running 100 million tokens per day on Claude Sonnet 4.6. On-demand pricing costs roughly $300 per day, while the same workload on a six-month provisioned commitment runs approximately $210 per day, saving $90 daily or $16,200 over the full term. However, you're locked in regardless of usage changes, so provisioned throughput makes sense only when your token volume is reliable within roughly 25 percent.

How to Choose Between Bedrock, Azure OpenAI, and Google Vertex AI

• Data Residency Requirements: If your workload requires data to remain within AWS infrastructure for compliance reasons, Bedrock is the only option among the three that guarantees data never leaves your VPC. Azure OpenAI and Google Vertex AI process data through their respective cloud environments.
• Model Diversity Needs: Bedrock offers the broadest model selection across all three platforms, with 15+ foundation models from multiple providers. Azure OpenAI focuses primarily on OpenAI models with some additional options, while Google Vertex AI emphasizes Google's Gemini family alongside third-party models.
• Cost Sensitivity at Scale: For high-volume, predictable workloads, Bedrock's provisioned throughput pricing can deliver 30 percent savings over on-demand rates. Azure and Google offer similar commitment options, but Bedrock's pricing on models like Nova Micro ($0.035 per million input tokens) is substantially lower than comparable alternatives on competing platforms.
• Enterprise RAG Workflows: Bedrock's integration of Cohere's Embed, Rerank, and Command R+ models creates a cohesive retrieval-augmented generation stack that delivers measurably better retrieval accuracy than mixing models from different providers, according to internal evaluations.
• Existing AWS Footprint: If your organization already runs infrastructure on AWS and uses IAM, KMS, and CloudTrail for governance, Bedrock integrates seamlessly without requiring separate vendor contracts or standing up new infrastructure.

The practical advantage of Bedrock's architectural design is that it turns frontier-model access into an AWS line item. For AWS-native enterprises that previously faced a choice between staying within their VPC and reaching cutting-edge models, that trade-off no longer exists. You get model access without leaving your AWS perimeter, without negotiating separate vendor contracts, and without managing your own inference infrastructure.

The expansion of Bedrock's model catalog in 2026 reflects a broader shift in how cloud providers compete on AI. Rather than building proprietary models exclusively, hyperscalers are positioning themselves as neutral platforms that aggregate the best models from multiple vendors. For enterprises, this means the decision to use Bedrock, Azure OpenAI, or Google Vertex AI increasingly comes down to existing cloud commitments, compliance requirements, and specific model preferences rather than artificial lock-in.