Quick Answer
As of Q4 2026, there are four ways to run Claude on AWS: Claude in Amazon Bedrock (the new Messages API integration for Opus 4.7 and later), the legacy Bedrock integration (Opus 4.6 and earlier), Claude Platform on AWS (Anthropic-operated, billed in Claude Consumption Units), and the direct Anthropic API. Per-token rates are identical on all four. Billing format, features, quotas, and cost visibility are not.
Here’s a line item CloudZero’s Claude Enterprise adapter was built to catch: a $1,200 user-level Claude charge landing against a Customer Success team’s $40,000 quarterly budget, with three weeks left in the quarter and the team pacing 18% over plan. Nobody chose to overspend. Someone just couldn’t see the pace.
Now put that same problem on AWS, where the plumbing multiplied twice in 2026. Claude Platform on AWS reached general availability on May 11, 2026. Then Amazon Bedrock itself split in two: a new Messages API integration for Claude Opus 4.7 and later, running beside the legacy InvokeModel and Converse APIs. Claude Opus 5.5 arrived on Bedrock on September 22, 2026 at $4 input and $20 output per million tokens.
Same models. Identical per-token rates. Four different bills, two quota systems, and feature gaps that quietly decide your unit economics. This guide maps every Claude on AWS path, what each one costs, and which one your workload actually needs.
What are the ways to run Claude on AWS?
Four surfaces serve Claude to AWS customers, and they split along one line: who operates the inference, AWS or Anthropic.
| Capability | Claude in Amazon Bedrock (new) | Claude on Bedrock (legacy) | Claude Platform on AWS | Direct Anthropic API |
|---|---|---|---|---|
| Operator | AWS (zero operator access) | AWS | Anthropic | Anthropic |
| API surface | Messages API (/anthropic/v1/messages) | InvokeModel, Converse | Full Claude API + Console | Full Claude API + Console |
| Models | Opus 4.7 and later (Fable 5.1, Opus 5.5, Sonnet 5, Haiku 4.5) | Opus 4.6 and earlier | All current models | All current models |
| Data boundary | Inside AWS | Inside AWS | Outside AWS (Anthropic) | Outside AWS (Anthropic) |
| Billing | AWS per-token line items | AWS per-token line items | AWS Marketplace, CCUs | Anthropic, per token |
| Auth | Bedrock service role, IAM, bearer tokens | AWS IAM + SigV4 | AWS IAM | Anthropic API key |
| Batch API (50% off) | No | Batch inference via S3, legacy models only | Yes | Yes |
| Web search, code execution | No | No | Yes | Yes |
| Managed Agents, MCP connector, Agent Skills | No | No | Yes (Bedrock has its own separate Agents) | Yes |
| Fast mode | No | No | No | Yes (Opus 5.5 at $8/$40) |
| New feature availability | Days to weeks after Anthropic | Legacy models only | Same day | Same day |
The row that decides most deployments isn’t a price. It’s the feature gaps. The new Claude AWS Bedrock integration keeps every token inside the AWS security boundary, with what Anthropic’s documentation calls “zero operator access” to the inference infrastructure. The trade: no Batch API, no server-side tools, no Managed Agents. Claude Platform on AWS delivers all of those on launch day, but your data processes on Anthropic’s side of the boundary.
For rates across every provider, not just Anthropic, our LLM API pricing comparison covers the field, and Anthropic’s consumer plans and full model rate card beyond AWS live in our Claude pricing guide.
2026 State of AI Spend Report
Why Finance Caps What It Can’t See
Learn how finance leaders are managing AI spend, and how they’re approaching FY2027
How much does Claude cost on Amazon Bedrock?
Bedrock matches Anthropic’s own rate card dollar for dollar on global endpoints. These are the current per-million-token rates and Bedrock model IDs:
| Model | Bedrock model ID | Input | Output | Cache read | Integration |
|---|---|---|---|---|---|
| Claude Fable 5.1 | anthropic.claude-fable-5-1 | $10 | $50 | $0.25 (0.025x) | New (Messages API) |
| Claude Opus 5.5 | anthropic.claude-opus-5-5 | $4 | $20 | $0.20 (0.05x) | New, added Sep 22, 2026 |
| Claude Opus 5 | anthropic.claude-opus-5 | $5 | $25 | $0.50 (0.1x) | New (Messages API) |
| Claude Opus 4.8 / 4.7 | anthropic.claude-opus-4-8 / -4-7 | $5 | $25 | $0.50 (0.1x) | New (Messages API) |
| Claude Sonnet 5 | anthropic.claude-sonnet-5 | $2 | $10 | $0.20 (0.1x) | New (Messages API) |
| Claude Haiku 4.5 | anthropic.claude-haiku-4-5 | $1 | $5 | $0.10 (0.1x) | New (Messages API) |
| Claude Sonnet 4.6 / 4.5 | ARN-versioned | $3 | $15 | $0.30 (0.1x) | Legacy (InvokeModel) |
| Claude Opus 4.1 | ARN-versioned | $15 | $75 | $1.50 (0.1x) | Legacy (InvokeModel) |
Older generations, including Claude Sonnet 4.5 on Amazon Bedrock and the Claude 3.5 family, remain available through the legacy integration at their original rates. Claude 2 is functionally retired from new workloads; if a 2023-era model is still in your routing config, that’s a migration ticket, not a pricing question.
Two things in that table should reset how finance teams read the Anthropic Bedrock lineup.
First, the pricing ladder inverted. Claude Opus 5.5 at $4/$20 costs 73% less than Opus 4.1 at $15/$75, two generations its senior. Teams still routing production traffic to older Opus models through the legacy integration are paying a multiple of the current frontier rate. That is a standing overcharge, not a discount opportunity.
Second, the current lineup prices like a doubling ladder: Haiku 4.5 at $1/$5, Sonnet 5 at $2/$10, Opus 5.5 at $4/$20. Each step up doubles the rate, which makes model routing math unusually clean.
One modifier sits on top everywhere: inference geography. On Bedrock it is endpoint choice, and the same 1.1x data-residency multiplier applies on Claude Platform on AWS and the first-party Claude API. Global endpoints, which route dynamically across all available regions, carry no premium. Regional endpoints, which guarantee routing through one geography for data residency, cost 10% more. The premium applies to every model from Claude Sonnet 4.5, Haiku 4.5, and Opus 4.5 onward, which covers the entire current lineup. Claude Opus 4.1 and earlier keep their original pricing structure. For how per-token billing works mechanically, our token-based pricing explainer breaks down the unit itself.
The full platform bill runs wider than Claude: Guardrails, Knowledge Bases, and AgentCore each carry their own meters, covered in our Amazon Bedrock pricing guide.
How does billing differ between Bedrock and Claude Platform on AWS?
Bedrock bills in tokens. Claude Platform on AWS bills in Claude Consumption Units. That single difference decides whether your AWS bill can answer “which team drove this month’s Claude spend” without extra tooling.
Bedrock billing: token line items
Claude Bedrock charges arrive as standard per-token line items, broken out by model. Cost allocation tags and AWS Cost Explorer work the way they do for any AWS service, and CloudTrail plus CloudWatch capture the activity trail. If your organization already runs tagging standards against Cost and Usage Reports, Claude on Bedrock spend drops into those workflows unchanged.
Claude Platform on AWS billing: CCUs
Claude Platform converts everything into CCUs before it reaches you:
- Anthropic rates your token usage in USD at standard per-model rates
- Negotiated discounts apply at conversion, metered as fewer CCUs
- The total converts at a fixed $0.01 per CCU (100 CCUs = $1.00)
- AWS Marketplace receives the CCU quantity hourly; invoices land monthly, arrears only
Your AWS bill shows one aggregated CCU line item. No per-model breakdown, no per-team split. The real-time breakdown lives in the Claude Console, reached through the AWS Console, while Cost Explorer sees only the aggregate. Procurement gets simplicity. Engineering attribution gets homework.
One warning from Anthropic’s own documentation: if you hold an existing Bedrock private offer, contact your account representative before activating Claude Platform. Discounts can’t be applied retroactively to usage that lands before the new private offer is accepted.
The same CCU mechanics now run on Azure through Claude in Microsoft Foundry, a pattern we track in our Azure OpenAI pricing guide. CCUs are becoming the marketplace-standard wrapper for Anthropic spend, and they aggregate away the very detail cost allocation depends on.
What does the same workload cost on each path?
Take an agent workload on Claude Opus 5.5: 20 million input tokens a month, 60% served from cache, plus 3 million output tokens. Ignoring one-time cache writes:
| Line item | Math | Cost |
|---|---|---|
| Uncached input | 8M x $4 per M | $32.00 |
| Cache reads | 12M x $0.20 per M | $2.40 |
| Output | 3M x $20 per M | $60.00 |
| Bedrock, global endpoint | $94.40 | |
| Bedrock, regional endpoint | $94.40 x 1.1 | $103.84 |
| Claude Platform on AWS | 9,440 CCUs x $0.01 | $94.40 |
Identical tokens, three different bill formats, and up to a 10% spread on price. The regional premium is the only case where Claude costs more on AWS than anywhere else. Without a data residency mandate, the global endpoint is the default and the regional premium is avoidable.
What are the AWS Bedrock Claude rate limits and quotas?
Claude in Amazon Bedrock starts every account at 2 million input tokens per minute (TPM). You can request up to 5 million input TPM and 500,000 output TPM without additional Anthropic approval, per Anthropic’s Bedrock documentation. Requests-per-minute limits are enforced separately on the AWS side, and RPM adjustments go through AWS support, not Anthropic.
Quotas are a cost-planning input, not just an ops detail, for two reasons.
First, throughput ceilings shape architecture. A workload that bursts past 2M input TPM either requests a raise, spreads across the global endpoint’s dynamic routing, or queues. The other two are capacity decisions with review cycles attached, so bake them into launch timelines rather than discovering them in an incident channel.
Second, the quota system replaces the commitment model for current models. Provisioned Throughput, Bedrock’s hourly reserved-capacity pricing, applies to the legacy integration. On the new Messages API integration, capacity is quota-governed and on-demand. Teams that budgeted around hourly commitments should re-model before migrating to Opus 5.5 or Sonnet 5.
Which cost optimizations work on each Claude deployment path?
The levers are the ones that govern all inference cost, but availability varies by path, and that changes the math.
- Route by task. The doubling ladder makes this concrete: moving classification and extraction from Sonnet 5 to Haiku 4.5 halves that traffic’s rate, and reserving Opus 5.5 for genuine agentic work caps the top of the curve. Routing works identically on every path.
- Cache aggressively, then verify the cache fires. Cache reads cost 10% of the input rate on most models, 5% on Opus 5.5, and 2.5% on Fable 5.1 and Mythos 5.1. On Bedrock, Anthropic flags that prompt caching may not be available in every region. If cache token counts sit at zero, you’re paying full input price for repeated context, and the bill won’t announce it.
- Batch what can wait, on the paths that allow it. The 50% batch discount is now a path decision. The Message Batches API runs on the direct API and Claude Platform on AWS. The new Bedrock integration doesn’t support it, and Bedrock’s S3-based batch inference runs on the legacy InvokeModel integration. A batch-heavy pipeline on Opus 5.5 has exactly two homes, and Bedrock isn’t one of them.
- Benchmark before migrating models. Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for identical text. Opus 5.5’s lower rate usually wins anyway, but run your real prompts through both models and compare total token counts before flipping production traffic.
- Normalize visibility across billing formats. Bedrock emits token line items. Claude Platform emits CCUs. Neither answers “what does Claude cost us per customer,” and the gap between those formats is where AI cost management programs go to die. More on closing it below.
Can you run Claude Code on AWS Bedrock?
Yes. Running Claude Code with Bedrock is officially supported, with Anthropic’s setup documentation covering the login wizard, IAM permissions, and credential refresh. It also runs through Claude Platform on AWS. Earlier in 2026 the honest answer was “only through Claude Platform,” and plenty of pages still say so. The docs moved. Most coverage didn’t.
What does Claude Code on Bedrock cost?
Sessions bill as ordinary Bedrock tokens on your AWS account, at the model rates in the table above. There’s no separate Claude Code AWS SKU: an agent session on Opus 5.5 meters at $4/$20 per million tokens like any other invocation, retiring against your AWS commitments, with IAM deciding who can run what.
That flips the procurement question. On Anthropic subscriptions, developer seats are a fixed monthly cost with usage limits, the model our Claude Code pricing guide breaks down plan by plan. On Bedrock, the same tool becomes metered AWS consumption: no per-seat ceiling, no rate-limit resets, and a bill that scales with exactly how much your engineers ship.
The Claude Code Bedrock setup carries three cost details worth flagging from Anthropic’s docs:
- Pin model versions for team deployments. New releases can silently shift which model your fleet calls, and with it your rate.
- The 1-hour cache TTL bills at 2x base input versus 1.25x for the 5-minute default. Long-running agent sessions should choose deliberately.
- Version floors apply to the newest models. Claude Code 2.1.255 or later runs Fable 5.1 on Bedrock; 2.1.280 or later runs Opus 5.5.
The desktop surfaces followed the CLI. Claude Desktop and Claude Cowork deploy against Bedrock through the Claude apps gateway, the self-hosted relay Anthropic documents for enterprise deployments, with corporate SSO and centralized policy enforcement.
CloudZero ships a Claude Code plugin that puts spend data inside the terminal where those sessions run, closing the loop between the engineer generating the cost and the number itself.
Which Claude on AWS path should you choose?
Choose Claude in Amazon Bedrock when data must stay inside the AWS boundary, when zero operator access is a compliance requirement, or when your cost allocation already runs on AWS tags and Cost and Usage Reports. Accept that batch discounts and server-side tools aren’t available there.
Choose Claude Platform on AWS when you build with Managed Agents, MCP connectors, or Agent Skills, when same-day model access matters, or when the Batch API’s 50% discount carries your unit economics. Accept CCU billing and data processing outside the AWS boundary. Managed Agents sessions add runtime billing at $0.08 per session-hour, metered to the millisecond while running, a meter Bedrock doesn’t have.
Choose the direct Anthropic API when you want one billing relationship and every feature, including fast mode, a research preview that runs Opus 5.5 at $8/$40 and exists on no AWS path at all. Access is gated through your Anthropic account manager.
The strongest argument for either AWS path is procurement, not features. Claude Platform usage billed through AWS Marketplace retires against existing AWS committed spend, and Bedrock consumption counts toward the same enterprise agreements. For organizations with large AWS commitments that’s real budget flexibility, the same reason enterprises route OpenAI spend through Azure instead of paying directly.
Most enterprises won’t choose one. Bedrock for regulated production inference, Claude Platform for agent development, a few direct API keys engineering spun up in Q1 that nobody decommissioned. That’s the realistic end state, and it’s a visibility problem before it’s a spending problem.
How does CloudZero unify Claude costs across every path?
Four deployment paths for one model family means Claude spend fragments across billing formats before anyone builds a unified view. Bedrock token line items in the AWS bill. A CCU aggregate from the Marketplace. A direct Anthropic invoice. And 87% of finance leaders are raising AI budgets they can’t justify, per our AI statistics roundup.
CloudZero was the first cloud cost platform with a direct Anthropic integration, and it ingests AWS billing data, Bedrock and Marketplace charges included, alongside it. Every source normalizes into cost per inference, allocated to the customer, feature, or team that drove it, whichever path served the tokens. You can see how the allocation engine handles it in the self-guided tour.
The payoff compounds at scale. Upstart, an AI lending platform, used CloudZero to help reduce cloud costs by $20 million. AI-native businesses don’t get there by reading rate cards. They get there by knowing, per customer and per feature, what the spend bought.
That’s the difference between monitoring and intelligence. Monitoring reports that the CCU line item grew. Intelligence reports which product’s agent traffic grew it, and whether the revenue attached to that product grew faster.
Schedule a demo to see your Claude spend, on every path it takes through AWS, in one view.
Frequently asked questions
