GILD verdict
Choose Bedrock as the main route when your AWS agreement, IAM and data boundary decide where models run, as for Zalando and Tines.
Run both if Claude carries production traffic: DoorDash and Fin keep Bedrock and Anthropic's API as two routes to the same models, so a quota cut or outage on one moves traffic to the other. Request quotas weeks ahead, since Ziva's never arrived. Choose Anthropic's API or Claude Platform on AWS when you need server tools, the MCP connector, the Files API, Opus 5.5 fast mode, or half-price batch on current models. At GILD's workload both routes cost US$3,264 a month with caching on Global routing; batch-ready work costs US$2,400 on Anthropic's API and has no batch option on Bedrock.
Teams at scale run both routes to the same Claude models. DoorDash keeps Bedrock as the fallback for its Claude API endpoint (InfoQ, Oct 3, 2026), and Fin serves Anthropic models through Bedrock, Vertex and Anthropic's own API (Aug 8, 2025). Teams that chose Bedrock alone, such as Zalando and Tines, did so for compliance and tenant data residency, while Ziva's Bedrock quotas sat 1,000 times below AWS's defaults and throttled up to 16% of requests (May 16, 2026). Anthropic's API wins on features, batch price and fast mode; Bedrock wins on procurement and the AWS data boundary. List prices match on Global routing, so at GILD's workload of 60,000 cached Opus 5.5 tasks a month both cost US$3,264.
What teams report
Failures and incidents 2
ZivaOn Amazon Bedrock
Ziva's Claude Haiku, Sonnet and Opus traffic on Bedrock ran into quotas 1,000 to 2,000 times below AWS's defaults (2,000 times for Opus), with 14% to 16% of requests throttled over 14 days, and four escalation channels over 16 days did not raise them (May 16, 2026).
“TL;DR: AWS is throttling our Bedrock usage at 1,000x below the minimum quota and won’t increase it.”
Ziva engineering, Company engineering blog, ZivaZiva blog, , write-up PositOn Amazon Bedrock
Claude Fable 5 on Posit Assistant's Bedrock provider returned HTTP 400 on every message, saying the default data retention mode was not available for the model; by Oct 23, 2026 the account-configuration cause and the retention modes the model accepts remained unconfirmed (GitHub issue, Jul 27, 2026).
“Selecting Claude Fable 5 on the Amazon Bedrock provider and sending any message returns HTTP 400 before generation…”
sharon-wang, Posit engineer, posit-dev/ai-lib maintainer, PositGitHub (posit-dev/ai-lib), , post
Run both 2
DoorDash
DoorDash's LLM Gateway calls the Claude API endpoint with Amazon Bedrock as its fallback, and OpenAI with Azure, so services stay up when one route fails; the gateway attributes cost to each department (InfoQ talk, Oct 3, 2026).
“We actually had to go understand what the fallback should be, what kind of routing should be. At one point, I remember we had OpenAI endpoint, but we also had a fallback on Azure endpoint. Or, we had Claude model API endpoint, then we had a fallback on Bedrock, so that the services can rely on us from the reliability point of view. We went in that direction. We quickly realized that as the usage is going up, then we need a way to track, because now finance got involved. Like, you guys have one API key for multiple teams, how do I track who is using what? We quickly realized we need workspaces where we have cost attribution. We can actually drill down on who is using how much. We can attribute cost back to those departments. That's what basically we did with LLM Gateway.”
Siddharth Kodwani, Software Engineer, GenAI Platform, DoorDashInfoQ, , recording Fin
Fin serves Anthropic models through AWS Bedrock, Google Vertex or Anthropic's own infrastructure, and moved traffic from OpenAI to Anthropic models during the OpenAI outage of June 10, 2025 (Aug 8, 2025).
“For example, Anthropic models can be served by AWS Bedrock, GCP Vertex, or Anthropic’s own infrastructure; OpenAI models are served by either Azure or OpenAI itself.”
Fin AI research team, Research team, FinFin AI Research, , write-up
Adopted 2
ZalandoOn Amazon Bedrock
Zalando's SRE team moved its postmortem analysis from small local models to Claude Sonnet 4 on Bedrock, driven mainly by compliance, and credits one finding with shielding it from 25% of later datastore incidents (Sep 25, 2025).
“…the current iteration is powered by Claude Sonnet 4 on AWS Bedrock. Such evolution was primarily driven by compliance topics…”
Dmitry Kolesnikov, Senior Principal Engineer, ZalandoZalando Engineering Blog, , write-up TinesOn Amazon Bedrock
Tines launched its first AI features on Bedrock to keep customer data in each tenant's Region and out of model training, and later kept a separate translation layer for Bedrock's format beside Anthropic and OpenAI-compatible clients (Mar 10, 2026).
“The first AI features we launched were built on AWS Bedrock in order to always maintain customer data within the same geographic regions as their tenant, as well as out of any LLM training pipelines.”
Yannick Gloster, Software Engineer, TinesTines blog, , write-up
Choosing between Amazon Bedrock and Claude API
Choose Amazon Bedrock if
- Your procurement and security approvals already cover AWS, and a new vendor contract would take months, as Zalando's legal review did for data with personal information.
- You want Claude next to OpenAI models, Amazon Nova, and open models under one IAM policy and one bill, or as the second route DoorDash and Fin keep for the same models.
- You need processing pinned to an AWS geography through Geo inference profiles, such as the EU or Japan.
Choose Claude API if
- You need the Files API, server-side tools such as web search and code execution, or the MCP connector on current models.
- Batch-heavy workloads: the Message Batches API halves Claude Opus 5.5 to US$2 input and US$10 output per million tokens.
- You need Opus 5.5 fast mode at US$8 input and US$40 output, which runs on the Claude API and not on Bedrock (Oct 28, 2026).
- You want new Claude features the day they ship; Claude Platform on AWS offers that too, billed through AWS Marketplace.
Side by side
| Dimension | Amazon Bedrock | Claude API |
|---|---|---|
| Monthly bill at one workloadGILD's calculation from published rates, fetched Oct 28, 2026: 2,000 tasks a day for 30 days, each with 2,000 new input tokens, 1,000 five-minute cache-write tokens, 7,000 cache-read tokens and 2,000 output tokens, at Opus 5.5 list prices of US$4 input, US$20 output, US$5 cache write and US$0.20 cache read per million tokens on both routes. The uncached lines send 10,000 input and 2,000 output tokens per task. Taxes and negotiated discounts are excluded. | US$3,264 a month on Claude Opus 5.5 with caching on Global routing; US$3,590.40 on Geo or In-Region routing. No batch option for Opus 5.5, so the uncached line is US$4,800415 | US$3,264 a month with caching; US$3,590.40 with US-only inference_geo. The same tasks uncached through the Message Batches API cost US$2,40026 |
| Fast mode | Not offered: the Opus 5.5 model card lists the Standard tier only15 | Opus 5.5 fast mode at US$8 input and US$40 output per million tokens, 2x the standard rate; not with the Batch API2630 |
| Older model retirement | Claude Sonnet 4.5 stays as Legacy until its end-of-life date of April 8, 202732 | Anthropic retires Claude Sonnet 4.5 on November 30, 202631 |
| Who runs it | AWS, on AWS-managed infrastructure with zero operator access; Anthropic staff have no access to it | Anthropic; Claude Platform on AWS is also Anthropic-operated and billed through AWS Marketplace |
| Claude Opus 5.5 list priceUS dollars per million input / output tokens. | US$4 / US$20 on Global cross-Region inference; US$4.40 / US$22 on Geo or In-Region | US$4 / US$20; US-only inference through inference_geo at 1.1x |
| Batch | Not listed for Claude Opus 5.5, Sonnet 5.5, Haiku 5.5, or Fable 5.1, and still absent from AWS's batch model table on Oct 25, 2026. Earlier models such as Opus 5 and Sonnet 4.6 get 50% off1415 | Message Batches API at 50% off on all current models |
| Cached input on Opus 5.5 | Writes US$5 (5-minute) or US$8 (1-hour); reads US$0.20 on Global routing | Writes 1.25x (5-minute) or 2x (1-hour) the input price; reads 0.05x |
| Features missing | The Files API and URL sources, server-side tools (code execution, web search, web fetch, advisor), Agent Skills, the MCP connector, Message Batches, Claude Managed Agents, and structured outputs on India inference profiles. Structured outputs work on every other profile across the Claude families, Opus 5.5 included15161724 | None on the Claude API; Claude Platform on AWS typically gets features the same day |
| Measured on a gateway | 57% of input tokens served from cache and 96% of requests completed on Requesty's gateway in April 202623 | 77% of input tokens served from cache and 96% of requests completed in the same data23 |
| CapacityZiva's new Bedrock account got 5 requests per minute for Opus 4.6 against a 10,000 default, and 16 days of escalation did not raise it (May 16, 2026). | Per-model, per-Region tokens-per-minute quotas; Anthropic lists a 2M input TPM default and up to 5M input and 500,000 output TPM without its approval, with RPM limits set by AWS2335 | Usage tiers with monthly spend caps: Start US$500, Build US$1,000, Scale US$200,000; Custom by arrangement28 |
| Data handling | AWS terms; inputs and outputs not shared with Anthropic; Fable 5 and 5.1 require aws_review, with up to 30 days of retention inside AWS | Anthropic terms; zero data retention by request per organization; Fable 5.1 and Mythos models require 30-day retention |
| Regions | Global routing, or Geo profiles that vary by model: US, EU, Australia, and Japan for Opus 5.5; US only for Fable 5.1 | Global by default; US-only inference available through inference_geo |
| Billing | AWS bill, with model charges listed under the provider through AWS Marketplace | Anthropic billing; Claude Platform on AWS bills through AWS Marketplace in Claude Consumption Units |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
Teams and leaders rate Claude API +25, 16 points above Amazon Bedrock at +9. Most of the gap is in complaints, with 20 of Amazon Bedrock's 57 sources negative against 9 of Claude API's 52.
The score counts the authors from teams and leaders that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so Claude API's +25 rests on votes split 20 positive, 20 mixed and 8 negative.
Each of Amazon Bedrock's authors wrote one source here, so each source is one vote. A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Run a pilot on your own prompts to measure output quality and cost per task.
Footprint on the AI Application Platform stack
They differ on six of 9 layers. Amazon Bedrock owns one that Claude API does not, and Claude API owns one that Amazon Bedrock does not. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Layer | Amazon Bedrock | Claude API |
|---|---|---|
| Model | ||
| Model gateway | Supports | Touches |
| Build | ||
| Tool use | Supports | Core |
| Evaluate | ||
| Evaluation | Supports | Touches |
| Operate | ||
| Observability | Supports | Touches |
| Cloud deployment | Core | Supports |
| Guardrails | Supports | Touches |
The 3 layers where they match
| Layer | Amazon Bedrock | Claude API |
|---|---|---|
| Model | ||
| Model APIs | Core | Core |
| Build | ||
| Orchestration | Supports | Supports |
| Data | ||
| Retrieval | Supports | Supports |
What practitioners praise
Among teams and leaders, Amazon Bedrock alone has three praise themes that reach 5 sources from 3 authors. Model Choice and AWS Integration lead, at 9 of 11 and 7 of 7. Both carry Cost Efficiency. It is praised in all 5 Amazon Bedrock sources that discuss it, against 8 of 8 for Claude API. Claude API alone has three praise themes that reach 5 sources from 3 authors. Model Quality and Tool Use & Structured Output lead, at 16 of 20 and 10 of 12.
| Praised for | Amazon Bedrock | Claude API |
|---|---|---|
| Cost Efficiency | 5 of 5 | 8 of 8 |
| Model Choice | 9 of 11 | Not a theme in its sources |
| AWS Integration | 7 of 7 | Not a theme in its sources |
| Managed Agents, RAG & Guardrails | 3 of 6 | Not a theme in its sources |
| Model Quality | 4 of 4 so far | 16 of 20 |
| Tool Use & Structured Output | 1 of 2 so far | 10 of 12 |
| Speed | 2 of 2 so far | 8 of 10 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among teams and leaders. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Among teams and leaders, Amazon Bedrock alone has five concern themes that reach 5 sources from 3 authors. Quotas, Capacity & Latency and Setup & Model Access lead, at 12 of 15 and 7 of 8. Claude API alone has five concern themes that reach 5 sources from 3 authors. Token Cost and Model Changes lead, at 11 of 12 and 8 of 11.
| Raised as a concern | Amazon Bedrock | Claude API |
|---|---|---|
| Quotas, Capacity & Latency | 12 of 15 | Not a theme in its sources |
| Setup & Model Access | 7 of 8 | Not a theme in its sources |
| Data Terms & Retention | 7 of 10 | Not a theme in its sources |
| Cost Surprises | 4 of 6 | Not a theme in its sources |
| Regional Availability | 2 of 5 | Not a theme in its sources |
| Token Cost | Not a theme in its sources | 11 of 12 |
| Model Changes | Not a theme in its sources | 8 of 11 |
| Tool Calling Failures | Not a theme in its sources | 5 of 7 |
| Latency | 4 of 4 so far | 3 of 5 |
| Setup & SDK Friction | Not a theme in its sources | 3 of 5 |
38 sources
- Amazon Bedrock User Guide: Supported Regions and models for batch inference
- Requesty: Operational metrics per provider, April 2026 (May 9, 2026)
- Requesty: Prompt-cache hit rate per provider, April 2026 (May 9, 2026)
- 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: Structured outputs for Anthropic Claude models
- langchain-aws pull request #1408: restore structured output for Claude 5 on Bedrock (Oct 20, 2026)
- Amazon Bedrock User Guide: Claude Fable 5.1 model card
- Amazon Bedrock User Guide: Anthropic models
- Amazon Bedrock User Guide: Batch inference
- Amazon Bedrock User Guide: Global cross-Region inference
- Amazon Bedrock User Guide: Data retention
- Amazon Bedrock User Guide: Quotas for the bedrock-runtime endpoint
- Claude API docs: Claude in Amazon Bedrock
- Claude API docs: Claude Platform on AWS
- Claude API docs: Pricing
- Claude API docs: Features overview
- Claude API docs: Rate limits
- Claude API docs: API and data retention
- Claude API docs: Fast mode
- Claude API docs: Model deprecations
- Amazon Bedrock User Guide: Claude Sonnet 4.5 model card
- InfoQ: Building a GenAI Platform at DoorDash, talk by Siddharth Kodwani and Swaroop Chitlur (Oct 3, 2026)
- Fin AI Research: Running a reliable service over unreliable parts (Aug 8, 2025)
- Ziva: 16 days, 4 channels, no AWS Bedrock quota fix (May 16, 2026)
- Zalando Engineering: AI-powered postmortem analysis (Sep 25, 2025)
- Tines: Futureproofing Tines, designing AI infrastructure for scale (Mar 10, 2026)
- posit-dev/ai-lib issue #35: Claude Fable 5 on Bedrock returns HTTP 400 (Jul 27, 2026)