Anthropic Direct vs AWS Bedrock vs Google Vertex AI in 2026

The three Claude API access paths compared in this guide. Vendor product pages linked above; brand marks omitted (publication policy: no approximation logos).

The bottom line

For a developer choosing how to call Claude in 2026, the question is rarely about model quality. The same Claude Sonnet 4.6 weights (the latest Sonnet release on Anthropic’s API since 17 February 2026, with Sonnet 4.5 still available alongside it; Opus 4.7 the latest frontier release as of 16 April 2026) are served by all three paths ^{1 3} . The choice is about per-token cost, billing rail, identity and access integration, and which cloud the rest of the stack already runs on.

Pick Anthropic direct when AWS or GCP is not already the team’s default cloud, when the lowest sticker price per token matters, or when fastest access to new model releases is non-negotiable. Anthropic ships new models on its own API first ³ , with Bedrock and Vertex following days to weeks later ¹ .

Pick AWS Bedrock when the application already runs on AWS and the team wants Claude under the same VPC, IAM, and CloudTrail surface as the rest of the workload ⁴ . Mumbai (ap-south-1) and Hyderabad (ap-south-2) are both enabled for Claude on Bedrock via Global cross-Region inference (CRIS), which routes inference traffic globally by default ⁵ . Teams that need data to stay inside a specific geography must explicitly pick a Geographic or In-Region inference profile ¹¹ .

Pick Google Vertex AI when the application is already on GCP and the team wants Claude alongside Gemini under one billing console, with Vertex’s MLOps stack (Pipelines, Model Registry, Endpoints) doing the heavy lifting ⁶ . Claude on Vertex is hosted in regional endpoints including asia-southeast1 (Singapore) and US/EU regions; asia-south1 (Mumbai) is not a Claude-on-Vertex region as of 2026-05-05 ⁷ . Singapore is the closest in-region option for workloads originating from the subcontinent.

The default for a developer paying out of pocket, with no AWS or GCP commit, is Anthropic direct. The other two earn their place when cloud lock-in is already a fact, not a choice. Microsoft Foundry is a fourth public-preview path on Azure carrying the same Claude lineup as Bedrock and Vertex (Sonnet 4.6, Opus 4.7, Haiku 4.5, plus older 4.x variants) and is flagged briefly later in this guide; the comparison architecture below covers the three production-grade paths ¹² . Pricing fluctuates across all three; verify each vendor’s pricing page before committing to a path.

How the comparison was built

The three paths in deep scope are the three GA production-grade ways to call Claude as of May 2026: Anthropic’s first-party API at api.anthropic.com, Amazon Bedrock’s managed Anthropic endpoint, and Google Vertex AI’s Model Garden Claude integration. Microsoft Foundry’s Claude integration is in public preview at writer-time and is summarised in a brief section below rather than folded into the full comparison until GA ¹² . OpenRouter, Together AI, and other aggregator routes are out of scope; they are useful for cross-model experimentation but not for production Claude workloads at scale.

Five factors carry weight in the recommendation. Per-token sticker price (input and output), at writer-time, including the regional-endpoint premium that applies to in-country routing on Bedrock and Vertex. Region availability, specifically what data-residency profiles each path offers for India-originating workloads. Authentication and identity integration with the rest of the cloud stack. Billing currency and invoice path (USD versus INR, GST handling). Time-to-availability of new Claude models on each path, since this is where the three diverge most visibly.

Two scope notes worth flagging. First, no head-to-head latency benchmark from India-routed endpoints exists across all three paths at writer-time; latency claims rely on each vendor’s published infrastructure footprint and region documentation. Second, the comparison covers the standard pay-as-you-go path on each provider, not Provisioned Throughput on Bedrock, committed-use discounts on Vertex, or Anthropic’s enterprise contracts, which all change the price math materially for high-volume workloads.

At a glance: the table

Axis	Anthropic Direct	AWS Bedrock	Google Vertex AI
Default Claude model (May 2026)	Claude Sonnet 4.6 (default since 17 Feb 2026); Sonnet 4.5 still available; Opus 4.7 latest frontier (16 April 2026)	Sonnet 4.6, Opus 4.7, Haiku 4.5 (parity; lag of days to weeks on new releases)	Sonnet 4.6, Opus 4.7, Haiku 4.5 (parity; lag of days to weeks on new releases)
Sonnet 4.6 input price (per 1M tokens, USD, global endpoint)	$3.00	$3.00 (Anthropic on-demand parity, global endpoint)	$3.00 (Anthropic on-demand parity, global endpoint)
Sonnet 4.6 output price (per 1M tokens, USD, global endpoint)	$15.00	$15.00 (Anthropic on-demand parity, global endpoint)	$15.00 (Anthropic on-demand parity, global endpoint)
Effective per-token premium vs Anthropic direct	Baseline	Global endpoint: matches Anthropic direct. Regional / multi-region endpoint (required to keep traffic in a specific geography): 10% premium over global	Global endpoint: matches Anthropic direct. Regional / multi-region endpoint (required to keep traffic in a specific geography): 10% premium over global
India-region routing	Region-routed by Anthropic; explicit India endpoint not published	Mumbai (ap-south-1) and Hyderabad (ap-south-2) enabled via Global cross-Region inference (default routes globally); Geographic / In-Region profiles available for in-country residency	asia-south1 (Mumbai) is NOT a Claude-on-Vertex region as of 2026-05-05; closest in-region option is asia-southeast1 (Singapore)
Authentication	API key (per workspace)	AWS IAM (signed requests, role-based access)	Google Cloud IAM (service accounts, OAuth)
Billing currency	USD only (international card)	USD on AWS India billing (depending on account); GSTIN supported on AWS invoices	USD on Google Cloud India billing (depending on account); GSTIN supported on GCP invoices
Time-to-new-model-availability	Day zero on Anthropic API	Days to weeks behind Anthropic direct	Days to weeks behind Anthropic direct
VPC / private networking	No (public endpoint)	Yes (VPC endpoints, PrivateLink)	Yes (Private Service Connect)
Best fit	Solo dev or team without AWS / GCP commit; lowest sticker; fastest access to new models	AWS-anchored team; needs IAM, VPC, CloudTrail audit logs; willing to pay 10% regional premium for in-country residency	GCP-anchored team; wants Claude alongside Gemini; accepts Singapore as nearest Claude-on-Vertex region

Anthropic Direct

Default Claude model (May 2026)

Claude Sonnet 4.6 (default since 17 Feb 2026); Sonnet 4.5 still available; Opus 4.7 latest frontier (16 April 2026)

Sonnet 4.6 input price (per 1M tokens, USD, global endpoint)

$3.00

Sonnet 4.6 output price (per 1M tokens, USD, global endpoint)

$15.00

Effective per-token premium vs Anthropic direct

Baseline

India-region routing

Region-routed by Anthropic; explicit India endpoint not published

Authentication

API key (per workspace)

Billing currency

USD only (international card)

Time-to-new-model-availability

Day zero on Anthropic API

VPC / private networking

No (public endpoint)

Best fit

Solo dev or team without AWS / GCP commit; lowest sticker; fastest access to new models

AWS Bedrock

Default Claude model (May 2026)

Sonnet 4.6, Opus 4.7, Haiku 4.5 (parity; lag of days to weeks on new releases)

Sonnet 4.6 input price (per 1M tokens, USD, global endpoint)

$3.00 (Anthropic on-demand parity, global endpoint)

Sonnet 4.6 output price (per 1M tokens, USD, global endpoint)

$15.00 (Anthropic on-demand parity, global endpoint)

Effective per-token premium vs Anthropic direct

Global endpoint: matches Anthropic direct. Regional / multi-region endpoint (required to keep traffic in a specific geography): 10% premium over global

India-region routing

Mumbai (ap-south-1) and Hyderabad (ap-south-2) enabled via Global cross-Region inference (default routes globally); Geographic / In-Region profiles available for in-country residency

Authentication

AWS IAM (signed requests, role-based access)

Billing currency

USD on AWS India billing (depending on account); GSTIN supported on AWS invoices

Time-to-new-model-availability

Days to weeks behind Anthropic direct

VPC / private networking

Yes (VPC endpoints, PrivateLink)

Best fit

AWS-anchored team; needs IAM, VPC, CloudTrail audit logs; willing to pay 10% regional premium for in-country residency

Google Vertex AI

Default Claude model (May 2026)

Sonnet 4.6, Opus 4.7, Haiku 4.5 (parity; lag of days to weeks on new releases)

Sonnet 4.6 input price (per 1M tokens, USD, global endpoint)

$3.00 (Anthropic on-demand parity, global endpoint)

Sonnet 4.6 output price (per 1M tokens, USD, global endpoint)

$15.00 (Anthropic on-demand parity, global endpoint)

Effective per-token premium vs Anthropic direct

Global endpoint: matches Anthropic direct. Regional / multi-region endpoint (required to keep traffic in a specific geography): 10% premium over global

India-region routing

asia-south1 (Mumbai) is NOT a Claude-on-Vertex region as of 2026-05-05; closest in-region option is asia-southeast1 (Singapore)

Authentication

Google Cloud IAM (service accounts, OAuth)

Billing currency

USD on Google Cloud India billing (depending on account); GSTIN supported on GCP invoices

Time-to-new-model-availability

Days to weeks behind Anthropic direct

VPC / private networking

Yes (Private Service Connect)

Best fit

GCP-anchored team; wants Claude alongside Gemini; accepts Singapore as nearest Claude-on-Vertex region

Anthropic direct: lowest sticker, fastest access, no cloud baggage

Anthropic’s first-party API at api.anthropic.com is the cleanest path to Claude. A workspace key, a POST request, a JSON response. Claude Sonnet 4.6 input is $3.00 per million tokens; output is $15.00 per million tokens, as of 2026-05-05; prices fluctuate, verify before committing ² . Those numbers are the reference price the cloud-routed paths benchmark against. Sonnet 4.5 remains available alongside Sonnet 4.6 at the same price band; Opus 4.7 (launched 16 April 2026) is Anthropic’s current frontier tier at higher per-token rates ² .

For a developer in India, the day-one friction is billing. Anthropic charges in USD via international card ² . There is no INR rail published, no UPI support, no Razorpay integration. The public GitHub issue requesting India-specific pricing remains open without a substantive Anthropic response on the thread ¹⁰ . Once the bank’s 2-3% forex markup and 18% GST on imported digital services are added at billing time, the effective bill lands meaningfully above the sticker. For a developer expecting a company-claim invoice in rupees with GSTIN, this matters; the workaround is a USD card with a personal expense, then claim through the company’s reimbursement flow.

What Anthropic direct earns is time-to-new-model. New Claude releases ship on api.anthropic.com first ³ . Sonnet 4.5 launched in September 2025 on Anthropic direct, Sonnet 4.6 in February 2026, and Opus 4.7 in April 2026; Bedrock and Vertex picked up each model in subsequent days to weeks per each platform’s documentation ¹ . For a team building against the frontier, the lag matters. For a team using a stable model in a production workload, the lag is invisible.

The other thing Anthropic direct earns is the lowest effective sticker for the use case article readers care about. On global endpoints, Bedrock and Vertex on-demand pricing for Sonnet 4.6 matches Anthropic direct ^{2 4} . On regional and multi-region endpoints, which are what an Indian workload needs to keep traffic inside a specified geography, Bedrock and Vertex apply a 10% premium over global ² . So the effective per-token cost for Mumbai-routed Bedrock or Singapore-routed Vertex is 10% above Anthropic direct’s global price, not at parity. The cloud premium also includes data egress, CloudWatch logging, Vertex Pipelines, and the opportunity cost of locking inference into one cloud’s billing line.

There is no VPC-native Anthropic endpoint. Traffic flows over the public internet between the developer’s compute and api.anthropic.com. For most workloads this is fine; for regulated workloads (financial services with RBI sensitivities, healthcare with sectoral data-residency rules, government use) the absence of private networking is the most likely reason to route through Bedrock or Vertex instead.

AWS Bedrock: the AWS-shop default

Amazon Bedrock is AWS’s managed-foundation-model service. Claude is one of several model families available; Anthropic and AWS have a deep commercial and infrastructure relationship that translates into broad Claude model coverage on Bedrock ⁴ . Sonnet 4.6, Opus 4.7, and Haiku 4.5 are available on Bedrock with the same on-demand pricing as Anthropic direct on global endpoints (Sonnet 4.6 at $3.00 input, $15.00 output per million tokens ² ). Regional and multi-region endpoints, used when traffic must stay inside a defined geography, carry a 10% premium over global ² .

Mumbai (ap-south-1) and Hyderabad (ap-south-2) are both enabled for Claude on Bedrock, but the access path matters. AWS delivers Claude in India via Global cross-Region inference (CRIS), where requests originating from ap-south-1 may be routed to inference endpoints anywhere globally to optimise availability ⁵ . By default, this means an India-originating Claude request on Bedrock does NOT keep its data inside India. To preserve in-country residency, teams must explicitly select a Geographic or In-Region inference profile when invoking the model ¹¹ . For Indian companies running customer-facing workloads with data-residency considerations, this distinction is load-bearing: the Geographic / In-Region path is what makes Bedrock-from-India a residency-preserving choice, and it carries the 10% regional-endpoint premium.

For an AWS-anchored team, the integration story is what wins. Bedrock requests are signed AWS IAM calls ⁴ . The same IAM role that grants access to S3 buckets, DynamoDB tables, and Lambda functions can grant fine-grained Bedrock-invocation permissions. Audit logs land in CloudTrail. Network traffic can flow over VPC endpoints with PrivateLink, never touching the public internet ⁴ . None of that is available on Anthropic direct.

The billing path is where the picture gets comfortable for companies on AWS India. AWS India bills through a local entity in many cases, supports GSTIN on invoices, and treats Bedrock charges as line items inside the existing AWS bill rather than as a separate vendor relationship. For procurement, security review, and finance team approval, “another AWS service” is meaningfully easier than “another vendor”. This is often the practical reason an AWS shop chooses Bedrock even when Anthropic direct’s sticker is identical.

Ancillary services add line items: CloudWatch logging on inference traffic, S3 storage of prompts and completions, data-egress charges if compute and Bedrock are in different regions. For heavily-instrumented production workloads these costs are real but small relative to the inference bill. The new-model lag is the other trade-off: new Claude releases reach Bedrock days to weeks after Anthropic direct, with the lag varying by model and AWS region ¹ . For a team locked to “ship the latest model the day it drops”, Bedrock is not the right path. For a team that values stability, audit posture, and integration over leading-edge model access, the lag is acceptable.

Image: AWS Bedrock product page (aws.amazon.com/bedrock), used for editorial coverage of the AWS-routed Claude API path compared in this guide.

Google Vertex AI: the GCP-shop default with the Gemini-plus-Claude story

Google Vertex AI is the Google Cloud equivalent of Bedrock: a managed-model platform that hosts Google’s first-party models (Gemini, Imagen, Codey) and a curated set of partner models in the Model Garden ⁶ . Claude is available in the Vertex Model Garden as a partner model ⁷ , with Sonnet 4.6, Opus 4.7, and Haiku 4.5 on the same Anthropic on-demand pricing of $3.00 input and $15.00 output per million tokens for Sonnet 4.6 on global endpoints, as of 2026-05-05; prices fluctuate, verify before committing ⁸ . Like Bedrock, regional and multi-region endpoints carry a 10% premium over global ^{2 9} .

Vertex’s region situation for Claude is a meaningful constraint for India-originating workloads. Claude on Vertex AI is hosted in regional endpoints us-east5, europe-west1, and asia-southeast1 (Singapore), with global and multi-region endpoints also available; asia-south1 (Mumbai) is NOT a Claude-on-Vertex region as of 2026-05-05 ⁷ . The asia-south1 region supports Vertex AI for Google’s first-party Gemini models, but the partner-model Claude integration is not available there. The closest in-region option for Claude on Vertex from India is asia-southeast1 (Singapore), which adds round-trip latency over what Bedrock’s Mumbai (ap-south-1) Geographic profile can offer. Verify Claude-specific region availability on the Vertex Model Garden documentation before architecting.

The Vertex-specific story is the Gemini-plus-Claude proposition. A team that wants to A/B test Claude against Gemini, route different request types to different models based on cost or capability, or use Gemini for cheap classification and Claude for high-stakes reasoning, can do all of that under one Vertex billing console, one set of GCP IAM service accounts, and one Vertex Pipeline. For applications where multi-model routing is a deliberate architectural choice rather than a hedge, Vertex is the natural home for that pattern.

For a GCP-anchored team, the integration story rhymes with Bedrock’s AWS story. Vertex requests are authenticated via Google Cloud IAM (service accounts, OAuth tokens) ⁶ . Audit logs flow into Google Cloud Audit Logs. Private Service Connect provides VPC-native networking. The MLOps stack around Vertex (Pipelines, Model Registry, Endpoints, Feature Store) is Google’s most-developed product line in the AI space and the reason a team on GCP would not seriously consider Anthropic direct as the production path.

The billing path on Google Cloud India is comparable to AWS India: GSTIN supported on GCP invoices, billing routed through Google’s local entity in many cases, and Vertex charges land as line items inside the existing GCP bill. The new-model lag is the same trade-off as Bedrock; Vertex picks up new Claude releases days to weeks after Anthropic direct ¹ . The Gemini side of the Vertex story is on day-zero parity with Google’s own model releases, so a team that values “latest Google model immediately” plus “Claude eventually” gets a reasonable best-of-both posture out of Vertex.

Image: Google Cloud Vertex AI product page (cloud.google.com/vertex-ai), used for editorial coverage of the GCP-routed Claude API path compared in this guide.

Use-case verdicts

Four developer profiles cover most Indian buying decisions on Claude API access paths in 2026. The pick for each is below; no path wins them all.

Solo developer or small team, no existing cloud commit

Winner: Anthropic direct. Runner-up: Bedrock if the team has already started on AWS for unrelated reasons.

The decision is about minimising friction. Anthropic direct gives the lowest sticker, the cleanest setup (key, request, response), and the fastest access to new models. The lack of INR billing is workable on a personal USD card. Spinning up an AWS or GCP account specifically to access Claude through Bedrock or Vertex adds operational overhead the team will not get back in value at small scale.

AWS-anchored team running customer-facing workloads from Mumbai

Winner: AWS Bedrock. Runner-up: Anthropic direct only if the team has not yet hit data-residency or audit constraints.

The integration story is decisive. IAM-signed requests, CloudTrail audit, VPC-native networking, and Mumbai (ap-south-1) plus Hyderabad (ap-south-2) availability via Bedrock together solve a basket of compliance and security questions an Anthropic direct integration cannot. The default Global cross-Region inference path routes traffic globally; teams that need data-residency posture must explicitly select the Geographic or In-Region inference profile and accept the 10% regional-endpoint premium ¹¹ . Procurement also goes faster when Claude shows up as another AWS service rather than a new vendor. The acceptable cost is days-to-weeks lag on new model releases.

GCP-anchored team running multi-model AI workloads

Winner: Google Vertex AI. Runner-up: Anthropic direct for one-off Claude experiments outside the production stack.

The Gemini-plus-Claude proposition is the win. A team that legitimately wants to route between models, A/B test, or compose pipelines that hit multiple models in sequence is better served by one Vertex Pipeline than by three separate vendor relationships. IAM, audit, and Private Service Connect cover the same compliance basket as Bedrock for AWS shops. The trade-off: Claude on Vertex is not in asia-south1 (Mumbai); the closest in-region path for Claude is asia-southeast1 (Singapore), which adds round-trip latency a Mumbai-resident workload would not pay on Bedrock ⁷ .

Enterprises with regulated data and in-country residency requirements

Winner: Bedrock with Geographic or In-Region inference profile, ahead of Vertex on the Mumbai-routing question (since Claude on Vertex is not in asia-south1 as of 2026-05-05). Runner-up: Vertex if the team accepts Singapore routing for Claude. Anthropic direct is structurally not the right path.

The deciding factors here are private networking, region routing, and an audit trail that satisfies internal security review and regulator expectations. RBI’s outsourcing guidelines for financial services, sectoral data-residency rules in healthcare, and the DPDP Act’s broader posture on cross-border data transfer all push toward “inference inside India, traffic inside the VPC, audit logs inside the cloud”. Bedrock with the Geographic or In-Region inference profile (10% regional-endpoint premium) satisfies the in-India routing requirement; Vertex’s nearest Claude region is Singapore, which does not. Anthropic direct does not satisfy the residency requirement at all.

Skip these specifically

Skip Anthropic direct for production workloads with Indian data-residency or audit requirements. The lack of VPC-native networking, the lack of in-country region routing, and the lack of an audit-log path that integrates with cloud-native tooling rule it out. The right answer is Bedrock or Vertex; do not engineer around Anthropic direct’s gaps when the cloud-routed path was built for exactly this use case.

Skip Bedrock for one-off Claude experiments outside an AWS commit. Setting up an AWS account, IAM roles, and Bedrock model access for a single proof-of-concept is operational overhead the experiment will not recover. Anthropic direct is the faster path to “is this idea worth pursuing”, with Bedrock onboarding deferred until the answer is yes and the production architecture is being designed.

Skip Vertex when the only AI model in the workload is Claude. Vertex’s value is the multi-model story plus the MLOps tooling around Gemini. A team using Claude alone, with no plans to add Gemini or other Vertex-hosted models, gets less differentiated value out of Vertex than a Gemini-shop or a multi-model team. Anthropic direct or Bedrock are both reasonable alternatives in that scenario.

Skip the cloud-routed path entirely if Anthropic announces an INR billing rail with GSTIN. The most common reason Indian solo developers and small teams default to Bedrock or Vertex today is the GSTIN invoice and AWS or GCP India billing relationship. If Anthropic ships the same on direct, the calculation flips for the small-scale segment overnight. The GitHub issue tracking the request is open ¹⁰ ; verify current state before betting either way.

What changes the calculation

Three things would shift the recommendation if they happen during 2026.

If Anthropic ships INR billing with GSTIN invoices for api.anthropic.com, the procurement-friendliness gap closes against Bedrock and Vertex for the small and medium segment. The capability case for Anthropic direct is already strong; the billing case is the only thing holding it back from being the default Indian-developer pick across more profiles.

If Google announces Claude availability in asia-south1 (Mumbai) on Vertex, the GCP-anchored profile gains a residency-preserving in-country path it does not have today. Bedrock already serves Claude from Mumbai (ap-south-1) and Hyderabad (ap-south-2) via Global cross-Region inference, with Geographic / In-Region profiles available; an additional Indian Vertex region for Claude would close the structural gap between AWS and GCP for regulated Indian workloads.

If Anthropic launches an enterprise tier with VPC peering or PrivateLink-equivalent connectivity directly on api.anthropic.com, the structural advantage Bedrock and Vertex hold for regulated workloads narrows. There is no public roadmap commitment to this; it is the most likely competitive move if Anthropic decides cloud-routed dependency is a strategic risk.

For now, model parity across all three paths means the model question is resolved. The cloud question is the live one. Pick the path that matches the rest of the stack, the billing rail the company already runs on, and the regulatory profile of the workload. Re-read this around late 2026, when the next round of cloud-side AI pricing announcements lands.