Key findings
- The table gives Standard on-demand examples in US dollars per 1M input and output tokens across eight providers. Read the exact row before budgeting.
- Batch is 50% below on-demand for supported models. Flex is 50% below Standard, Priority is 75% above Standard, and Reserved tier or Provisioned Throughput require model-specific capacity checks.
- Claude in Amazon Bedrock lacks the Files API, Anthropic's server-side tools, the MCP connector, and Managed Agents. Structured outputs now work across the Claude model families, and a langchain-aws contributor found Bedrock still rejects the schema on India inference profiles.
- Data terms changed in 2026. Claude Fable 5 and 5.1 now require the aws_review mode, which keeps traffic inside AWS for up to 30 days, and AWS states that Bedrock shares nothing with model providers.
- Of 149 public accounts on Bedrock, 29 say procurement or data control drove the choice, 28 discuss quotas, capacity, or latency, and 12 discuss models or features arriving late.
Amazon Bedrock sells models from many providers. Standard on-demand prices vary by model, and Batch, Flex, Priority, Reserved tier, Provisioned Throughput, cross-Region inference, and prompt caching change each model's price or capacity. Teams choose it for AWS procurement, identity, Regions, and data controls. Check the exact model row and feature support before routing traffic.
Representative Bedrock model prices
AWS publishes a separate price row for each model and tier. The table below shows Standard on-demand prices in US dollars per 1M input and output tokens across the providers a buyer can pick on Bedrock. These examples use US East (N. Virginia), except the Claude row, which uses Global routing. Bedrock has no single platform rate.
AWS also lists older models, such as Llama 2 Chat (13B) at US$0.75 / US$1.00 and Jurassic-2 Mid at US$12.50 / US$12.50 per 1M input / output tokens. The table leaves them out and shows current model families only.
The model row sets the price. Batch, Flex, Priority, cross-Region routing, prompt-cache reads and writes, Reserved tier, and Provisioned Throughput each have their own support and rates.
| Model and provider | Input / 1M tokens (US$) | Output / 1M tokens (US$) | Output price per 1M tokens |
|---|---|---|---|
| Anthropic Claude Opus 5.5 (Global) | 4.00 | 20.00 | |
| Amazon Nova Pro | 0.80 | 3.20 | |
| Meta Llama 4 Scout 17B | 0.17 | 0.66 | |
| Mistral Large 3 | 0.50 | 1.50 | |
| DeepSeek V3.2 | 0.62 | 1.85 | |
| OpenAI gpt-oss-120b | 0.15 | 0.60 | |
| Cohere Command R | 0.50 | 1.50 | |
| AI21 Jamba 1.5 Mini | 0.20 | 0.40 |
AWS Amazon Bedrock pricing page, checked October 26, 2026; rates read from AWS's published price list and names cross-checked against each AWS model card.
Features that stay with the vendor
Anthropic lists these Bedrock gaps: the Files API and URL sources, server tools such as code execution and web search, Agent Skills, the MCP connector, Message Batches, and Claude Managed Agents. Structured outputs have left that list. AWS documents them across the Claude Sonnet, Opus, Haiku, Fable, and Mythos families, and its Opus 5.5 card lists them on bedrock-runtime and bedrock-mantle. The same card shows no token counting on bedrock-runtime, and the model supports only the Standard tier. A langchain-aws contributor testing every Region profile on October 20, 2026 found Bedrock still rejects the schema on India inference profiles.
Practitioners report delayed models and missing API features. A Claude Code team using a Bedrock proxy waited for models such as Opus 4.5. A builder found that Converse rejected the documented 1-hour cache. Other accounts report months of delay for open models.
Quotas and Regions
Bedrock quotas apply per model and Region in tokens per minute. Ziva's quotas were 1,000 to 2,000 times below AWS's documented defaults. Four escalation channels over 16 days failed to raise them. Another team's Opus 4.7 quota fell to zero overnight while it served government customers.
Geo profiles keep processing inside a geography such as the EU, a stated reason several teams use Bedrock. A German health platform uses Claude in eu-central-1 because legal review ruled out Anthropic's API for health data. Regional availability also delays adoption. Sonnet 5 was missing from the EU at launch.
| Concern | Discuss it | Raise it as a problem | Share of sources that discuss it |
|---|---|---|---|
| Quotas, capacity & latency | 28 | 23 | |
| Setup & model access | 16 | 15 | |
| Regional availability | 13 | 8 | |
| Lag behind vendor APIs | 12 | 10 | |
| Cost surprises | 11 | 9 | |
| Data terms & retention | 11 | 8 |
GILD analysis of 149 public accounts, Feb 2025 to Oct 2026, opened Oct 12 to Oct 28, 2026
Data terms after the Fable change
In June and July 2026, practitioners reported a provider data-sharing requirement for Anthropic's newest Bedrock models. Teams that had promised customers their data stayed in AWS objected. AWS now documents a different arrangement. Claude Fable 5 and 5.1 require aws_review, which keeps prompts and outputs inside AWS for up to 30 days. The legacy provider_data_share mode shares nothing today.
Retention is set per Region at the account or project level, and an organization can enforce zero retention with a Service Control Policy, which makes models that require review unavailable to it.
Why teams accept the gaps
Procurement is the most common reason for choosing Bedrock. A Fortune 500 non-tech company buys Anthropic models through a click-through EULA. A multi-billion-dollar AWS customer still needed 4 to 6 months of legal review before using Claude and OpenAI agents. Consultants report clients that buy only through Bedrock or Azure.
Methodology
GILD built this piece from AWS's Amazon Bedrock pricing page, model cards, and data retention, quota, and service tier documentation. It also used Anthropic's pricing and Claude in Amazon Bedrock pages. All were checked October 12, 2026. AWS's structured outputs page for Claude and the Opus 5.5 model card were rechecked October 25, 2026. AWS pricing, its regional rate feeds, and the eight representative model cards were rechecked October 26, 2026. Prices are standard list prices in US dollars per million tokens. Every price, quota, and feature figure is AWS's or Anthropic's own published number. Account counts come from 149 public accounts published between February 5, 2025 and October 27, 2026, the same set behind the Amazon Bedrock Signals. Each account cited here is linked in Sources. Speed figures come from those accounts, including an Artificial Analysis benchmark of Claude Sonnet 5 across four providers.
Products in this piece
Sources
- Official documentation
- 21
- Public source
- 13
- Amazon Bedrock pricing
- AWS model card: Anthropic Claude Opus 5.5
- AWS model card: Amazon Nova Pro
- AWS model card: Meta Llama 4 Scout 17B
- AWS model card: Mistral Large 3
- AWS model card: DeepSeek V3.2
- AWS model card: OpenAI gpt-oss-120b
- AWS model card: Cohere Command R
- AWS model card: AI21 Jamba 1.5 Mini
- AWS price list: metered model rates
- AWS price list: foundation model rates
- Amazon Bedrock User Guide: Claude Opus 5.5 model card
- Amazon Bedrock User Guide: Anthropic models
- Amazon Bedrock User Guide: Global cross-Region inference
- Amazon Bedrock User Guide: Data retention
- Amazon Bedrock User Guide: Abuse detection
- Amazon Bedrock User Guide: Quotas for the bedrock-runtime endpoint
- Claude API docs: Claude in Amazon Bedrock
- Claude API docs: Pricing
- Claude API docs: Features overview
- Amazon Bedrock User Guide: Structured outputs for Anthropic Claude models
- langchain-aws pull request #1408: restore structured output for Claude 5 on Bedrock (Oct 20, 2026)
- Ziva: 16 days, 4 channels, no AWS Bedrock quota fix (May 16, 2026)
- Hacker News: Claude Opus 4.7 quota suddenly changed to 0 TPM in Bedrock (May 1, 2026)
- Hacker News comment on Claude Code through a Bedrock proxy (Feb 6, 2026)
- Hacker News comment on Converse API cache points (Apr 30, 2026)
- Hacker News comment on Bedrock trailing new open models (Aug 17, 2026)
- Hacker News comment on Sonnet 5 missing from Bedrock's EU region (Jun 30, 2026)
- MedPolaris on GitHub: Configure Claude via AWS Bedrock eu-central-1 for EU data residency (Apr 14, 2026)
- Hacker News comment on keeping data in-country with Bedrock (Jun 2, 2026)
- Hacker News comment on Opus 5 and provider data sharing on Bedrock (Jul 24, 2026)
- Hacker News comment on vetting Bedrock models for data retention (Jun 10, 2026)
- Hacker News comment on 4 to 6 months of legal review for Bedrock (Sep 10, 2026)
- Hacker News comment on large corporate deployments requiring Bedrock (Jun 1, 2026)