Amazon Bedrock Review (2026): Features, Pricing & Verdict

What Is Amazon Bedrock?

Amazon Bedrock is AWS’s fully managed foundation-model service — the layer that turns the world’s largest cloud into a generative-AI platform — and its defining idea has held steady since launch while everything around it scaled: Bedrock is not a model, it’s a control plane. The model you call is a configuration decision — change a model ID and you’ve switched providers — while the security, networking, observability and billing layer stays constant and stays AWS. What has transformed is the catalogue behind that API. From a curated handful at launch, Bedrock in mid-2026 fronts more than 100 foundation models from 18-plus providers, and the roster reads like a peace treaty: Anthropic’s full Claude line (the platform’s most popular family, through the Claude 4 generation, with the newest Opus-class model bringing a 1M-token context and 128K output to Bedrock for the first time); OpenAI’s frontier models — GPT-5.5, GPT-5.4, Codex and the open-weight gpt-oss line — added through the April 2026 partnership expansion that ended the era of OpenAI-means-Azure exclusivity; Amazon’s own Nova 2 family spanning Micro/Lite/Pro/Premier text tiers, Canvas (image), Reel (video) and Sonic (speech-to-speech, GA early 2026); plus Meta Llama, DeepSeek-R1, Mistral, Cohere, AI21, Stability AI, Google Gemma, Qwen, Writer, Luma AI, TwelveLabs and NVIDIA Nemotron, with a Marketplace extending into specialised and emerging models beyond the core catalogue. Around the models, the platform stack is what enterprises actually buy: Bedrock Agents and the newer AgentCore for autonomous multi-step workflows; Knowledge Bases for managed RAG (bring your data, Bedrock handles chunking, embedding and retrieval); Guardrails for content filtering and hallucination checks at $0.15 per thousand text units — 80% cheaper than at launch; Flows for visual workflow orchestration; Data Automation for turning unstructured documents, audio and video into structured output; model evaluation (automatic, or human at $0.21 per task); fine-tuning, reinforcement fine-tuning and distillation; custom model import (bring your own weights, pay only per model-copy serving time); and Intelligent Prompt Routing, which routes each request between model tiers and can cut costs up to 30% without accuracy loss. Billing comes in modes that reward planning: on-demand per-token (the default), batch at 50% off for asynchronous jobs, provisioned throughput for reserved capacity on 1- or 6-month commitments, and prompt caching that can cut input costs up to 90% on repetitive-context workloads. All of it runs inside the AWS trust perimeter — VPC isolation, IAM, CloudTrail, the compliance certification stack, 44 regions including GovCloud, and cross-region inference billed at source-region rates with no routing surcharge. Within our Model Providers & AI Infrastructure category, Bedrock is the aggregation thesis at enterprise scale: where every other review in this series asks “is this provider’s model worth adopting,” Bedrock asks “why adopt a provider at all when you can adopt them all behind one door” — and for a large share of the enterprise market, that question has already been answered.

Core Features

The catalogue: one API, every frontier

Bedrock’s first-order feature is breadth with neutrality, and 2026 is the year the breadth became genuinely complete. The strategic milestone was April’s OpenAI partnership expansion: for the first time, an enterprise can call Anthropic’s Claude Opus-class models and OpenAI’s GPT-5.5 through the same API, the same IAM policies, the same VPC endpoints and the same bill — which converts the industry’s defining rivalry into a dropdown menu. The practical consequences are larger than they sound. Model migration collapses from a re-architecture project to a configuration change: the unified Converse API means switching from Claude to Nova to Llama is an ID swap, and Intelligent Prompt Routing automates the choice per-request (routing simple queries to cheap models, complex ones to frontier models, with documented savings up to 30% at equivalent accuracy). Vendor negotiation leverage inverts: teams that once locked into one provider’s pricing now benchmark continuously across the catalogue, and Bedrock’s price parity with going direct (Claude Sonnet-class at the same ~$3/$15 you’d pay Anthropic; in-region OpenAI inference at parity with OpenAI’s data-residency tier) means aggregation costs nothing at the token level. Coverage extends across modalities — text and code from every major family, DeepSeek-R1 for open reasoning at high-volume prices, Stability and Nova Canvas for image, Nova Reel and Luma for video, Nova Sonic for speech-to-speech, TwelveLabs for video understanding, plus embeddings and rerank models for the RAG stack — and across trust tiers, from open weights (gpt-oss, Llama, Gemma, Qwen) you could later self-host, to frontier closed models under enterprise terms. The first-party Nova 2 family anchors the value floor: Micro at $0.035 per million input tokens is one of the cheapest capable models anywhere — the routing/classification/extraction workhorse — with Lite around $0.06, Pro at $0.80/$3.20 handling mid-tier work, and Premier at the top; because they’re first-party, Nova models are consistently the platform’s price leaders, and for the large fraction of enterprise workloads that are structured extraction and classification rather than frontier reasoning, they carry the load at near-negligible cost. The honest limits of the catalogue thesis: the newest releases still reach their home platforms first (a new Claude or GPT typically lands on Anthropic/OpenAI direct days-to-weeks before Bedrock, and regional rollout across 44 regions is uneven), a handful of notable providers remain absent or partial, and the sheer menu — 79-plus models with published pricing spanning $0.04 to $18.80 per million input tokens — shifts the burden of model selection onto the buyer, which is a real cost the single-provider path doesn’t carry.

The enterprise stack: agents, RAG, guardrails and governance

The catalogue gets the headlines; the platform services around it are why enterprises standardise on Bedrock rather than a thin router. The agent layer leads in 2026: Bedrock Agents configure autonomous assistants that connect to data sources, maintain session memory, call tools and APIs, and decompose tasks into steps — with AgentCore, the 2026 addition, providing the production-grade runtime (identity, memory, observability, tool gateways) that turns agent prototypes into governed deployments; neither carries a separate platform fee, with billing flowing from the model calls, retrievals and guardrail checks the agent triggers — an elegant structure whose flip side is that multi-step reasoning multiplies token consumption in ways that dominate real agentic bills. Knowledge Bases deliver managed RAG end-to-end — point them at your documents and Bedrock handles chunking, embedding, vector storage and retrieval, with rerank models (billed per query, 100 chunks per query unit) sharpening relevance — and the critical 2026 buying note is the vector-store choice underneath: OpenSearch Serverless, the historical default, carries a ~$345/month two-OCU floor even at zero traffic, while S3 Vectors (launched December 2025) runs up to 90% cheaper, supports trillions of vectors at sub-second latency, and is now the right default for most workloads. Guardrails provide model-agnostic safety — content filters, denied topics, PII redaction, contextual-grounding checks against hallucination — at $0.15 per thousand text units after an 80% price cut, with no charge for blocked requests; because they’re a platform layer, one policy applies identically across Claude, GPT, Nova and Llama, which is the governance property compliance teams actually want. The remaining services fill out the lifecycle: Flows for visual orchestration of models, agents and AWS services; Data Automation for industrialised document/audio/video-to-structured-data processing (the intelligent-document-processing workhorse); model evaluation with algorithmic scores free beyond inference and human evaluation at $0.21 per task; fine-tuning and the newer reinforcement fine-tuning (Bedrock automates the RFT loop against your reward function, billed hourly, with the tuned model served on-demand after); distillation for shrinking frontier behaviour into cheap models; and custom model import, which serves your own fine-tuned open weights inside Bedrock’s managed environment billed per model-copy in five-minute increments — no charge for the import itself. Wrapped around everything is the property no standalone provider can replicate: the AWS operational fabric — IAM-scoped access per model per team, CloudTrail audit on every invocation, CloudWatch token/latency/spend metrics, VPC endpoints keeping traffic off the public internet, and the compliance certification stack (HIPAA, SOC, FedRAMP-adjacent GovCloud) that lets regulated industries deploy generative AI without a new security review per vendor. This stack is Bedrock’s real moat: models are increasingly commodities; governed, observable, multi-model production infrastructure is not.

Billing modes, optimisation levers and the true cost of ownership

Bedrock’s billing is simultaneously the most flexible in this category and the most demanding of operator skill — the gap between a tuned deployment and a naive one routinely exceeds the gap between providers. The modes: on-demand (per-token, no commitment, the default — with cross-region inference at source-region pricing giving capacity resilience for free); batch at a flat 50% discount for asynchronous jobs (the single easiest saving for any workload that tolerates latency — evaluations, backfills, bulk document processing); provisioned throughput, which works like reserved instances — model units guaranteeing throughput, billed hourly on 1- or 6-month terms, economical above roughly 60–70% utilisation and a money-loser below; and prompt caching, which discounts repeated context by up to 90% and is the highest-leverage optimisation for agent and RAG workloads whose prompts share large static prefixes — documented as dramatically underused. Layer the routing tools on top — Intelligent Prompt Routing between model tiers (up to 30% savings at maintained accuracy), and simple model right-sizing across the catalogue (the difference between reflexively calling a frontier model and routing classification to Nova Micro is literally two orders of magnitude: $0.035 versus multi-dollar input rates) — and disciplined teams can drive unit costs down relentlessly. Now the honest ledger of where undisciplined budgets die, because the documented pattern is consistent: real bills land 1.5–2x initial estimates, and the causes are structural, not hidden. Output-token asymmetry — across nearly every model, output costs 3–5x input (Claude Sonnet-class at $3/$15 being the canonical example), so chat-heavy and verbose-reasoning workloads cost far more than input-based estimates suggest. Agent amplification — a single user request through a multi-step agent generates many model calls, retrievals and guardrail evaluations, each billed; agentic production is where “$100/month prototype” becomes “$5,000/month platform.” Adjacent services — the OpenSearch floor under Knowledge Bases, CloudWatch metrics and logs, data transfer ($0.09/GB out to internet; free within-region with VPC endpoints), evaluation-judge tokens billed at standard rates — accrue beside the Bedrock line item and outside most estimates. And customisation carries its own meter: fine-tuning and RFT bill training hours, distillation bills the teacher model’s synthetic-data generation at on-demand rates, custom model copies bill serving minutes whether traffic arrives or not. None of this is predatory — every rate is published — but Bedrock is a platform whose economics reward engineering attention and punish set-and-forget, and the right mental model is the one AWS veterans already hold: the bill is a design artefact.

Scored Categories

Pricing

Frequently Asked Questions