Claude API vs Bedrock vs Vertex AI: Pricing and Limits

If you want Claude in production, you can pay Anthropic directly, go through Amazon Bedrock, or go through Google Cloud's Vertex AI, which Google renamed Gemini Enterprise Agent Platform in April 2026. At the default global endpoints, the per-token list prices for current Claude models are the same on all three. What differs is which features you get, who sets your rate limits, who processes your data, and whose invoice the charge lands on.

This guide compares those things using each vendor's own documentation and pricing pages, and ends with a plain recommendation for the most common situations.

There are more than three ways to buy Claude

Anthropic's documentation now lists several distinct routes, and the names matter because they behave differently:

• Claude API (first party). You sign up with Anthropic, use API keys and the Claude Console, and Anthropic bills you. New features and betas land here first.
• Claude in Amazon Bedrock. AWS operates the inference stack. It now serves Claude through a Messages API endpoint at /anthropic/v1/messages, while the older InvokeModel and Converse integration remains for Opus 4.6 and earlier models. Usage is billed as an AWS service. See our Amazon Bedrock page.
• Claude Platform on AWS. Anthropic operates the service, but you sign in with AWS IAM and pay through AWS Marketplace. It became generally available on May 11, 2026. See Claude Platform on AWS.
• Claude on Google Cloud, Google managed. This is what most people still call Claude on Vertex AI. Google serves the models through Model Garden on Gemini Enterprise Agent Platform, and the charge shows up as Agent Platform usage.
• Claude on Google Cloud, Anthropic managed. Anthropic runs the Claude API inside your Google Cloud project and bills through Google Cloud Marketplace. We covered that launch in Google Cloud now lets Anthropic run Claude.

Microsoft Foundry is a further option for Azure shops, but this comparison sticks to the three that searchers usually weigh against each other: Anthropic direct, Bedrock and Vertex AI.

Token prices: identical at the global endpoint

Prices as of Oct 27, 2026, in US dollars per million tokens, from the Claude pricing page, the Amazon Bedrock pricing page (read through the AWS Price List API for us-east-1, Standard tier) and Google's Agent Platform pricing page. At the global endpoint, all three list the same input and output rates:

• Claude Opus 5.5: $4 input and $20 output on all three.
• Claude Sonnet 5.5: $2 input and $10 output on all three.
• Claude Haiku 5.5: $0.10 input and $0.50 output for prompts up to 100K tokens, and $0.50 and $2.50 above that, on all three.
• Claude Sonnet 5: $2 input and $10 output on the Claude API and on Bedrock's global endpoint.

Prompt caching follows the same pattern. On the Claude API, a cache hit costs $0.20 per million tokens for Opus 5.5 and $0.10 for Sonnet 5.5, and Bedrock's global cache-read rates match those figures. A one-hour cache write for Opus 5.5 is $8 on both. AWS notes on its pricing page that it cut the Sonnet 5.5 cache-read price by 50% from Oct 7, 2026.

The 10% premium for keeping data in one place

The price gap appears when you need data residency. Bedrock's regional cross-region inference profiles (prefixes such as us., eu. or jp.) and Google's multi-region and regional endpoints both cost 10% more than global for Claude Sonnet 4.5 and later. On Bedrock in us-east-1, that makes Sonnet 5.5 $2.20 input and $11 output, Opus 5.5 $4.40 and $22, and Haiku 5.5 $0.11 and $0.55. Google lists Sonnet 5.5 at $2.20 and $11 on its regional and multi-region endpoints.

Anthropic has its own version of this. Setting inference_geo: "us" on the Claude API, Claude Platform on AWS or the Anthropic-managed Google Cloud offering applies a 1.1x multiplier. Either way, pinning traffic to a geography costs about a tenth more, so only pay it when compliance actually requires it.

Batch, tools and other extras

• The Claude API's Message Batches endpoint halves token prices, for example Opus 5.5 at $2 input and $10 output.Bedrock lists batch inference at 50% off on-demand for select models, including Anthropic's. Bedrock also has Standard, Flex, Priority and Reserved service tiers. Its page only states the Flex and Priority adjustments for other providers' models, so check your Claude model before planning around them.Google's pricing page has separate batch lines for Claude. Neither Bedrock nor the Google-managed offering supports Anthropic's own Message Batches API, so batch jobs go through each cloud's batch service instead.Server-side tools are priced by Anthropic: web search costs $10 per 1,000 searches, web fetch has no extra charge, and code execution includes 1,550 free hours a month and then costs $0.05 an hour. On Claude in Amazon Bedrock, none of those server-side tools is available.Opus 5.5 fast mode costs $8 input and $40 output. Anthropic lists it for its own API, the Anthropic-managed Google Cloud offering and Foundry's Azure-hosted deployments. Bedrock and the Google-managed Vertex route aren't on that list.

One more cost note that applies everywhere: Anthropic says the tokenizer used from Claude 4.7 onward produces roughly 30% more tokens for the same text. If you're migrating from an older model, re-measure token counts before comparing bills.

Features: where the cloud routes lag

This is the biggest practical difference. Anthropic's docs list what each partner route doesn't support:

• Claude in Amazon Bedrock doesn't support URL sources for images and documents, the Files API, server-side tools (code execution, web search, web fetch, advisor), Agent Skills, the MCP connector, programmatic tool calling, Claude Managed Agents, server-side fallbacks, or the Batches, Models, Admin, Compliance and Usage and Cost endpoints. It does support prompt caching, extended thinking, tool use including computer use, citations and structured outputs.
• Claude on Google Cloud (Google managed) has a very similar gap list: no URL sources or Files API, no code execution, web fetch or advisor, no Agent Skills, MCP connector or programmatic tool calling, no Managed Agents, no server-side fallbacks and none of those admin endpoints. It supports prompt caching and the 1M-token context window on current models.
• Claude Platform on AWS is the exception on AWS. Anthropic says its features typically arrive the same day as on the Claude API, and the AWS launch post lists Managed Agents, web search and fetch, code execution, the Files API, Skills, the MCP connector and batch processing.

Feature timing also differs. Anthropic describes Bedrock's feature availability as following the Amazon Bedrock release schedule, and both clouds set their own model lifecycle and retirement dates, which can differ from Anthropic's. If you need a new beta the week it ships, the first-party API or one of the Anthropic-operated cloud routes is the only safe bet.

Rate limits and quotas

Who sets your limits depends on who operates the service.

• Claude API: Anthropic uses usage tiers. On the Start tier, Opus 5.5, Sonnet 5.5 and Haiku 5.5 each allow 1,000 requests, 2,000,000 input tokens and 400,000 output tokens per minute. Monthly spend caps are $500 on Start, $1,000 on Build and $200,000 on Scale, and the Custom tier has none. Cache reads don't count toward the input-token limit on most models, so good caching raises your effective throughput.
• Claude Platform on AWS: the same tier system applies. New organizations start on Start and move up as they build a history of paid AWS Marketplace invoices. Anthropic says it uses a separate capacity pool from both the Claude API and Bedrock, so you can fail over between them.
• Amazon Bedrock: AWS manages rate limits and quotas, and you request increases through AWS. Model access has to be enabled in the Bedrock console, and some models are open to everyone while others have their own access criteria. Anthropic's Bedrock guide also notes a 20 MB request payload limit.
• Google Cloud (Google managed): Google Cloud manages quotas, and you raise them through Google's quota page for Anthropic models. Agent Platform caps request payloads at 30 MB. Regional endpoints serve Claude Sonnet 4.6 and earlier models only, so newer models need the global or multi-region endpoint.

For spiky workloads, the useful takeaway is that limits are separate per route. Teams that hit ceilings sometimes split traffic across the Claude API and a cloud route, which is the same idea behind the gateways in our AI gateways comparison.

Data, security and billing

• Bedrock runs on AWS-managed infrastructure with what Anthropic calls zero operator access, meaning Anthropic staff can't reach the inference stack. AWS is the data processor. Anthropic itself points regulated teams that need FedRAMP High, IL4, IL5 or HIPAA-ready setups, or AWS as the only processor, to Bedrock.
• Claude Platform on AWS keeps Anthropic as the data processor under Anthropic's data terms. AWS says requests are processed outside the AWS security boundary. It supports AWS PrivateLink, and zero data retention is available on request. Billing goes through AWS Marketplace and, per Anthropic, counts against existing AWS commitments.
• Google managed puts data handling under Google Cloud's terms and bills Claude as Agent Platform usage. Access uses Agent Platform IAM roles such as roles/aiplatform.user.
• The Claude API is billed by Anthropic directly, typically through credit purchases, and uses Anthropic API keys and workspaces rather than cloud IAM.

If your company already has a committed spend agreement with AWS or Google, routing Claude through that cloud can matter more to finance than any per-token difference, because the list prices are the same.

What switching looks like in code

Anthropic's SDKs cover every route, so switching mostly means changing the client and the model name. The following names come from Anthropic's docs and are illustrative, not a complete setup:

• Claude API: the standard Anthropic client with an API key, model ID claude-sonnet-5-5.
• Claude in Amazon Bedrock: pip install -U "anthropic[bedrock]" and the AnthropicBedrockMantle client, endpoint https://bedrock-mantle.{region}.api.aws/anthropic/v1/messages, model ID anthropic.claude-sonnet-5-5.
• Claude Platform on AWS: the AnthropicAWS client (in beta), base URL aws-external-anthropic.{region}.api.aws.
• Google managed: pip install -U "anthropic[vertex]" and the AnthropicVertex client with your project ID and region="global". Requests go to :rawPredict or :streamRawPredict with the model named in the URL.

Anthropic's ant command-line tool doesn't support Bedrock or the Google-managed offering, so scripts built around it need the SDK instead.

When to pick each

• Pick the Claude API if you're a startup or small team without a big cloud commitment, you want new features and betas on day one, or you rely on server-side tools, Agent Skills, the Files API or Managed Agents.
• Pick Claude in Amazon Bedrock if you need AWS to be the only data processor, you work under FedRAMP, HIPAA-ready or similar requirements, or your stack already lives on AWS and you can live without server-side tools. Bedrock is also the natural home if you use other Bedrock features, such as the ones in our piece on Bedrock web search domain filtering.
• Pick Claude Platform on AWS if you want the full Anthropic feature set but need AWS sign-in, CloudTrail audit logs and one AWS invoice, and you don't need AWS to process the data.
• Pick Claude on Vertex AI (Google managed) if your workloads and spend commitments are on Google Cloud and you want Claude next to Gemini in Model Garden under Google's quotas and terms. If you need Anthropic's newest features on Google Cloud, look at the Anthropic-managed offering instead.

Verdict

Don't choose on token price, because at the global endpoint there isn't a difference. Choose on three questions: who must process your data, which features you need, and which bill your company prefers. For most independent developers and small teams, the Claude API is the simplest and most complete. For regulated AWS shops, Bedrock is the safer default. For Google Cloud shops, the Google-managed Vertex route keeps everything inside one console. Whichever you pick, use the global endpoint unless compliance forces a region, since the regional options add 10%.

Sources

• Anthropic: Claude pricing
• Amazon Bedrock pricing
• Google Cloud: Agent Platform pricing (Vertex AI)
• Anthropic: Claude in Amazon Bedrock
• Anthropic: Claude on Amazon Bedrock (Opus 4.6 and earlier)
• Anthropic: Claude Platform on AWS
• Anthropic: Claude Models on Google Cloud overview
• Anthropic: Claude on Google Cloud (Google managed)
• Anthropic: Rate limits
• AWS: Claude Platform on AWS is now generally available
• Google Cloud: Gemini Enterprise Agent Platform name changes