Claude Code: Bedrock vs. Claude Platform vs. Direct
A practical architecture, governance, and cost guide for engineering leaders evaluating Claude Code consumption paths in 2026.
The short answer
There are three primary ways to use Claude and Claude Code when AWS is part of your technology strategy:
- Anthropic Direct — Anthropic operates the service, provides the API or subscription, and bills you directly.
- Amazon Bedrock — AWS operates the model-serving environment. You access Claude using AWS authentication, APIs, networking, logging, and billing.
- Claude Platform on AWS — Anthropic operates the service while AWS provides the authentication and billing entry point through your AWS account.
Published base token rates are broadly aligned across these paths, but their effective cost and operating model are not identical. Subscriptions, negotiated discounts, inference geography, provisioned capacity, feature charges, logging, networking, and governance can all change the total cost.
The more durable question is therefore not simply, “Which endpoint is cheaper?” It is:
Who should operate inference, where may data be processed, which product features do we need, and how do we want to govern access and spend?
Understanding the three paths
Dimension | Anthropic Direct | Amazon Bedrock | Claude Platform on AWS |
|---|---|---|---|
Service operator | Anthropic | AWS | Anthropic |
Primary API | Anthropic Messages API | Bedrock Invoke/Converse APIs or supported Bedrock-native endpoints | Anthropic Messages API |
Authentication | Anthropic credentials or subscription sign-in | AWS IAM/SigV4 or supported Bedrock authentication | AWS IAM/SigV4 or API key |
Billing | Anthropic | AWS service bill | AWS bill using Marketplace metering |
Data processor for model service | Anthropic | AWS | Anthropic and AWS |
Feature release timing | Native Anthropic availability | Depends on AWS integration and regional availability | Generally same-day with the first-party Claude API |
Best fit | Simplicity or non-AWS operating model | AWS-controlled inference and AWS-native governance | Native Claude API features with AWS authentication and billing |
AWS’s official comparison confirms that Bedrock is AWS-operated while Claude Platform on AWS is Anthropic-operated. It also explains that Claude Platform uses the native Messages API and generally receives Anthropic features on the first-party release timeline. See Claude Platform on AWS vs. Amazon Bedrock .
Pricing: similar list rates, different economics
As of August 5, 2026, Anthropic publishes the following API list prices per million tokens:
Model | Input | Output | Cache write* | Cache read |
|---|---|---|---|---|
Claude Haiku 4.5 | $1 | $5 | $1.25 | $0.10 |
Claude Sonnet 5** | $2 | $10 | $2.50 | $0.20 |
Claude Opus 5 | $5 | $25 | $6.25 | $0.50 |
Claude Fable 5 | $10 | $50 | $12.50 | $1.00 |
*Cache-write prices shown are for the five-minute TTL. Extended caching has separate rates.
**Sonnet 5 promotional pricing of $2/$10 per million input/output tokens runs through August 31, 2026; published standard pricing is $3/$15 afterward.
Amazon Bedrock lists the same Sonnet 5 promotional rate, and Claude Platform on AWS applies Anthropic’s per-model and per-feature rates before converting the charge into Claude Consumption Units (CCUs). A CCU is fixed at $0.01, but a negotiated discount can reduce how many CCUs are metered. See Claude Platform billing .
That is why “the price is identical everywhere” is too simple. Total economics may differ because of:
- Anthropic Pro, Max, Team, or Enterprise subscriptions versus metered API consumption
- Private offers and negotiated discounts
- US-only, geographic, global, or in-Region inference choices
- Batch processing, prompt caching, service tiers, and provisioned capacity
- Separate charges for tools such as web search, code execution, or managed-agent runtime
- The infrastructure and operations required for gateways, networking, telemetry, and governance
Subscriptions are not token bundles
Claude Code is included in Anthropic’s paid plans. Current published prices include Pro at $20 monthly, Max starting at $100 monthly, Team Standard at $25 per seat monthly, and Team Premium at $125 per seat monthly, with annual discounts available for Pro and Team. Enterprise starts at $20 per seat plus usage billed according to model and task.
It is tempting to convert those prices into a token “break-even” point, but that comparison is unreliable. Subscription limits use rolling session and weekly windows shared across Claude surfaces. Actual capacity depends on the model, context length, conversation complexity, and features used; there is no fixed message or token entitlement.
For planning, run a representative pilot. Anthropic reports that enterprise Claude Code API deployments average approximately $150–$250 per active developer per month, with substantial variation. See Manage Claude Code costs .
Security and data-processing boundaries
The strongest architectural distinction is who operates inference.
With Amazon Bedrock, AWS is the service operator and data processor. AWS documents that model providers cannot access Bedrock deployment accounts, prompts, completions, or service logs. Bedrock can be reached privately from a VPC using AWS PrivateLink, and customer-controlled logging destinations can use AWS Key Management Service encryption.
This does not mean the model runs inside your VPC or AWS account. Bedrock runs on AWS-controlled service infrastructure. Processing location also depends on the inference option you select:
- In-Region inference processes in a supported Region.
- Geographic cross-Region inference can route within a defined geography such as the US or EU.
- Global cross-Region inference can route among supported commercial Regions worldwide.
AWS states that cross-Region traffic remains on the AWS network and is encrypted in transit. Organizations with residency requirements should choose an eligible in-Region or geographic option rather than assuming every Bedrock request remains in one Region. See Bedrock data protection and cross-Region inference .
With Claude Platform on AWS, Anthropic processes inference outside the AWS security boundary. AWS IAM still governs access to Claude Platform resources, and billing flows through AWS, but those controls do not turn the service into AWS-operated inference.
With Anthropic Direct, identity, service operation, billing, and data processing are managed through Anthropic’s platform and commercial agreement.
Audit, observability, and cost attribution
Bedrock offers the deepest AWS-native integration, but all three paths have useful controls.
Amazon Bedrock
Bedrock integrates with CloudTrail for AWS API activity and can publish model invocation logs to CloudWatch Logs or Amazon S3. Invocation logging can include request and response bodies, token counts, model ID, request ID, and caller identity.
Two qualifications matter:
- Model invocation logging is disabled by default and must be configured in each applicable Region.
- It currently covers supported operations through the
bedrock-runtimeendpoint; other endpoints may not be captured.
For billing attribution, application inference profiles, Bedrock projects, workspaces, and IAM principal attribution can attach costs to teams or workloads. These produce aggregated billing records in Cost Explorer and Cost and Usage Reports—not a bill line for every prompt. Per-prompt analysis requires invocation logs and optional request metadata. See Bedrock model invocation logging and Bedrock cost-attribution guidance .
Claude Platform on AWS
Claude Platform operations can be audited through CloudTrail. Workspace and vault management operations are logged as management events; inference, batch, file, skill, model, user-profile, and agent activity are data events that require explicit CloudTrail data-event configuration and may add CloudTrail charges.
Claude Platform does not currently publish per-workspace token and latency metrics as native CloudWatch metrics. Usage and model breakdowns are available through the Claude Console and Anthropic Usage and Cost API. Workspace tags can flow to CUR for cost allocation after the tag keys are activated. See Claude Platform monitoring and logging .
Anthropic Direct
Anthropic’s business plans provide usage reporting and spend controls. Enterprise offerings add capabilities such as audit logs, role-based access, custom retention controls, analytics, and compliance APIs. These are Anthropic-native controls rather than CloudTrail or CloudWatch integrations.
The balanced conclusion is that Bedrock provides the strongest AWS-native invocation-content logging and telemetry integration, not that the other paths provide no auditing or observability.
Resiliency and capacity
Bedrock cross-Region inference can improve throughput by selecting model capacity across eligible Regions. This is capacity routing, not a complete disaster-recovery design. If your source endpoint, application Region, or dependencies fail, you may still need application- or gateway-level multi-Region failover.
Provisioned Throughput reserves a defined level of model throughput at a fixed hourly cost. It can improve capacity predictability for sustained workloads, but availability, model-unit sizing, commitment terms, and economics vary by model. It should not be described as universally eliminating throttling or guaranteeing application latency. Cross-Region inference profiles also do not currently support Provisioned Throughput.
Claude Platform and Anthropic Direct manage rate limits through Anthropic. Higher limits may be requested, but they do not provide the same Bedrock resource model for reserved throughput.
For enterprise Claude Code fleets, the Claude apps gateway for AWS adds a useful control plane. The self-hosted gateway provides OIDC sign-in, centralized policy, OTLP telemetry, per-user or group spend caps, and routing to Bedrock or Claude Platform on AWS. It can also be configured for cross-Region or cross-account upstream failover.
Feature availability: choose the product surface you need
Claude Platform on AWS is designed to provide the native Claude API experience with AWS authentication and billing. Current capabilities include prompt caching, batch processing, extended thinking, computer use, Files API, code execution, Agent Skills, MCP connectors, and Claude Managed Agents. Some features remain unavailable or limited, including HIPAA readiness, certain admin APIs, some managed-agent capabilities, and workspace-level default inference-geography controls. The authoritative list is Claude Platform feature support .
Bedrock provides a different set of AWS-native building blocks:
- Bedrock Guardrails
- Knowledge Bases for managed retrieval-augmented generation
- AgentCore services and runtime capabilities
- Cross-Region inference profiles
- Provisioned Throughput for supported models
- AWS-native identity, logging, networking, and cost attribution
- Access to models from multiple providers through a common AWS service
These capabilities are not always one-for-one substitutes. For example, AgentCore Code Interpreter and Anthropic’s Messages API code-execution tool solve related problems through different APIs and operational models.
Model and feature availability also varies by Region, endpoint, release stage, and integration timeline. Avoid assuming that every Claude model or beta feature is simultaneously available on every path.
Compliance: eligibility is not automatic compliance
AWS states that Claude Platform on AWS is a third-party offering and is not covered by standard AWS compliance programs or audit reports. Organizations must assess Anthropic’s controls and agreements separately.
Bedrock is covered by multiple AWS compliance programs, with eligibility depending on the Region, feature, and workload configuration. Organizations should verify the current AWS services-in-scope documentation and their AWS agreements rather than treating a checkmark as proof that a complete workload is compliant.
Anthropic Direct also has enterprise compliance capabilities. Anthropic offers HIPAA-ready Enterprise and first-party API configurations under a Business Associate Agreement, but coverage is feature-specific. Certain Claude Code surfaces require zero-data-retention configuration, and some beta or tool features are excluded.
A more accurate decision rule is:
- Choose Bedrock when you require AWS-operated inference or controls covered by applicable AWS compliance programs.
- Evaluate Anthropic Direct when Anthropic’s enterprise controls, BAA, retention settings, and eligible product surfaces meet the workload requirements.
- Do not use Claude Platform on AWS for HIPAA workloads under its current documented feature set.
Compliance is ultimately a property of the full architecture, configuration, contracts, data flow, and operating process—not the model name alone.
AWS commitments and private pricing
Anthropic’s Claude Platform on AWS announcement states that its AWS billing can retire existing AWS commitments. Bedrock and Marketplace treatment, discount eligibility, private-offer terms, and any limits are contract-specific.
Do not assume that a general EDP (Enterprise Discount Program) percentage automatically discounts third-party model tokens, or that every contract treats Marketplace and Bedrock consumption identically. Confirm the following before making a financial recommendation:
- Which spend retires the commitment
- Whether retirement is full or subject to contractual limits
- Whether a private offer must be accepted first
- Whether discounts apply to token rates, feature charges, or only the metered quantity
- Whether treatment differs between Bedrock and Claude Platform on AWS
A practical decision framework
Choose Amazon Bedrock when:
- AWS must operate inference and act as the model-service data processor.
- You need AWS-native IAM, PrivateLink, CloudTrail, invocation logging, or cost allocation.
- You need an eligible AWS compliance scope or AWS-controlled processing boundary.
- You want Guardrails, Knowledge Bases, AgentCore, or multi-provider model access.
- Geographic routing or reserved throughput is important.
Choose Claude Platform on AWS when:
- You want the native Claude Messages API and same-day Claude feature access.
- You want AWS IAM authentication and AWS billing.
- Anthropic-operated processing outside the AWS boundary is acceptable.
- Its documented compliance posture and current feature limitations fit the workload.
Choose Anthropic Direct when:
- You prefer the simplest direct relationship with Anthropic.
- AWS-native identity, billing, and observability are not requirements.
- Subscription economics are attractive for individual or team interactive use.
- Anthropic’s enterprise controls and contractual terms fit your organization.
A hybrid pattern can be reasonable: use Bedrock for workloads that require AWS-operated inference and governance, Claude Platform for selected native API features, and subscriptions for eligible interactive developer workflows. The trade-off is additional platform complexity, so establish clear workload-routing and data-classification rules.
Cost optimization that applies across paths
Whatever path you choose, focus on measurable optimization rather than universal savings percentages:
- Select the least expensive model that meets the quality target. Reserve high-capability models for tasks that benefit from them.
- Use prompt caching for repeated context. Include cache-write cost and TTL behavior in the calculation.
- Use batch processing for supported asynchronous workloads. Confirm feature and model availability first.
- Keep context intentional. Clear unrelated sessions, compact carefully, and inspect what consumes the context window.
- Control extended thinking. Reasoning tokens are billed as output; use the capability where its quality benefit justifies the cost.
- Measure with a pilot. Track cost per active developer, task, accepted change, or business outcome—not just tokens.
Conclusion
The three paths can look similar on a pricing page, but they represent different operating models.
- Bedrock emphasizes AWS-operated inference, AWS-native governance, and broader AWS platform integration.
- Claude Platform on AWS emphasizes native Claude API capabilities with AWS authentication and billing.
- Anthropic Direct emphasizes a direct product and commercial relationship with Anthropic, including subscription options.
Start with data-processing and compliance requirements, then evaluate API features, operational controls, capacity, and commercial terms. Base token prices matter, but they should not decide the architecture by themselves.
The right question is not “Which Claude endpoint is cheapest?” It is:
Which operating boundary gives this workload the right combination of capability, control, and cost?
Primary sources
Pricing note: Prices and product capabilities change frequently. Treat this article as a decision framework, not a procurement quote. Verify current rates on the Amazon Bedrock pricing page and Anthropic pricing page , and confirm private pricing and commitment treatment with your account teams.
The views expressed here are my own and do not represent the official position of any organization. Product availability, compliance eligibility, prices, and contractual treatment can change; verify them against current official documentation before making architecture or procurement decisions.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article