# Claude on third-party cloud platforms — Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS

Status: DOCUMENTED (no cloud credentials — nothing live-tested)

Sources (all retrieved 2026-09-18, canonical URLs):

• https://platform.claude.com/docs/en/build-with-claude/claude-in-amazon-bedrock
• https://platform.claude.com/docs/en/build-with-claude/claude-on-amazon-bedrock-legacy
• https://platform.claude.com/docs/en/build-with-claude/claude-platform-on-aws
• https://platform.claude.com/docs/en/api/claude-platform-on-aws-iam-actions
• https://platform.claude.com/docs/en/build-with-claude/claude-on-vertex-ai
• https://platform.claude.com/docs/en/build-with-claude/claude-in-microsoft-foundry
• https://platform.claude.com/docs/en/build-with-claude/overview (Feature availability tables)
• https://platform.claude.com/docs/en/models/overview and models/*/overview pages, https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions, https://platform.claude.com/docs/en/about-claude/model-deprecations
• https://platform.claude.com/docs/en/about-claude/pricing, https://platform.claude.com/docs/en/manage-claude/data-residency, https://platform.claude.com/docs/en/api/supported-regions, https://platform.claude.com/docs/en/api/rate-limits, https://platform.claude.com/docs/en/api/beta-headers, https://platform.claude.com/docs/en/api/overview
• Feature guides under build-with-claude/* and agents-and-tools/** (platform notes), cli-sdks-libraries/sdks/{python,typescript}
• Machine-readable twins: generated/fragments/compatibility/anthropic-feature-platform-matrix.json, generated/fragments/compatibility/anthropic-model-platform-ids.json, generated/fragments/endpoints/anthropic-cloud-endpoints.json

Last verified: 2026-09-18 (docs only; every record carries verification.result = "restricted", request_note = "no cloud credentials")

Other platforms checked: grep -rli 'snowflake|databricks|oracle|cloudflare|salesforce' over sources/anthropic/pages returns no platform pages (only unrelated hits in api/errors.md, api/compliance.md, MCP-tunnels pages). The Anthropic docs document exactly five API platforms.

# 1. Platform overview

Is Claude Platform on AWS a distinct endpoint? Yes. It is a separate product from Bedrock: Anthropic-operated infrastructure reached through an AWS-owned gateway (aws-external-anthropic.{region}.api.aws), billed by AWS Marketplace, IAM-scoped per workspace (ARN arn:aws:aws-external-anthropic:{region}:{account-id}:workspace/wrkspc_…), with a separate capacity pool from both the first-party API and Bedrock. Signing up provisions a new Anthropic organization (no conversion of an existing org). Prerequisite: aws iam enable-outbound-web-identity-federation once per account (otherwise every request fails with "Outbound web identity federation is disabled for your account"). Same-day feature parity with the Claude API except the items in §6. All AWS commercial regions; PrivateLink supported. Workspaces are bound to one region; the region controls the gateway, not where inference runs (use inference_geo).

# 2. Per-platform details

# 2.1 Claude in Amazon Bedrock (Messages-API endpoint, "bedrock-mantle")

• POST https://bedrock-mantle.{region}.api.aws/anthropic/v1/messages; curl: --aws-sigv4 "aws:amz:{region}:bedrock-mantle". Standard Anthropic client also works with base_url=https://bedrock-mantle.{region}.api.aws/anthropic + bearer token as api_key (bearer only; SigV4 needs the dedicated client).
• Models: anthropic.claude-fable-5-1, anthropic.claude-fable-5, anthropic.claude-opus-5 (access criteria), anthropic.claude-opus-4-8, anthropic.claude-opus-4-7, anthropic.claude-sonnet-5, anthropic.claude-haiku-4-5, anthropic.claude-mythos-preview (invite, dedicated allowlisted account, regional us-east-1 only).
• Endpoint types: Global (dynamic routing, no premium) and Regional (+10 %, data residency; inference profiles US / EU / JP / AU). 27 regions listed (af-south-1 … us-west-2); "In-region only" = direct single-region routing without inference profile: ap-northeast-1, ap-southeast-4, eu-north-1, eu-west-1, us-east-1, us-east-2, us-west-2. Fable 5.1 regional = us-east-1 only.
• Supported highlights: Messages API, prompt caching, thinking, tool use (bash, computer use beta versions, memory, text editor), citations. Not supported: structured outputs (see contradiction in §8), URL/Files input sources, server-side tools (code execution, web search, web fetch, advisor), Agent Skills, MCP connector, programmatic tool calling, Message Batches / Models / Admin / Compliance / Usage-Cost endpoints, Managed Agents, server-side fallbacks, computer_toolset_20260801 / browser_toolset_20260801.
• Logs: CloudWatch + CloudTrail; support bedrock-ant-eap@amazon.com with request-id.

# 2.2 Amazon Bedrock legacy integration (InvokeModel / Converse, Opus 4.6 and earlier)

• ARN-versioned ids; newer models (Fable 5.1/5, Opus 5/4.8/4.7, Sonnet 5) are also reachable through InvokeModel on bedrock-runtime (served by the Mantle infrastructure) but have no ARN-versioned id.
• Passing a base id fails with HTTP 400 "…with on-demand throughput isn't supported. Retry your request with the ID or ARN of an inference profile…" → use an inference-profile id (global., us., eu., jp., apac. prefix) or ARN arn:aws:bedrock:{region}:{account-id}:inference-profile/{id}.

• Body: Messages API minus model, plus "anthropic_version": "bedrock-2023-05-31". Streaming = AWS EventStream. Global vs regional (CRIS) pricing (+10 %) applies from Sonnet 4.5 onward.
• Legacy-only differences: no automatic prompt caching (top-level cache_control → 400; use explicit breakpoints); structured outputs are available for Opus 4.6 / Sonnet 4.6 / Sonnet 4.5 / Opus 4.5 / Haiku 4.5; PDF via Converse needs citations for visual analysis (InvokeModel for full control); tool search via InvokeModel only; mid-conversation system messages only via InvokeModel for Mantle-served models; 1M context on Opus 4.6 / Sonnet 4.6, 200k on Sonnet 4.5 and older.
• List models: aws bedrock list-foundation-models --by-provider anthropic.

# 2.3 Claude Platform on AWS (Anthropic-operated)

• POST https://aws-external-anthropic.{region}.api.aws/v1/messages with --aws-sigv4 "aws:amz:{region}:aws-external-anthropic" (region in URL must match the SigV4 region) + anthropic-version + anthropic-workspace-id. Env: AWS_REGION (required, no default), ANTHROPIC_AWS_WORKSPACE_ID, ANTHROPIC_AWS_API_KEY. Credential precedence: apiKey arg → awsAccessKey+secret → awsProfile → ANTHROPIC_AWS_API_KEY → default AWS chain.
• Models (Claude API ids, no prefixes): claude-fable-5-1, claude-fable-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-5, claude-sonnet-4-6, claude-sonnet-4-5, claude-haiku-4-5. Mythos models are not listed.
• IAM: namespace aws-external-anthropic, 71 actions, VerbNoun (e.g. CreateInference → POST /v1/messages, CountTokens, CreateBatchInference/GetBatchInference (metadata + results)/ListBatchInferences/CancelBatchInference/DeleteBatchInference, GetModel/ListModels, CreateFile/GetFile/ListFiles/DeleteFile, workspace actions, ListComplianceActivities, AssumeConsole, CallWithBearerToken (route-less, needed for API-key auth)). Managed policies: AnthropicFullAccess, AnthropicReadOnlyAccess, AnthropicInferenceAccess, AnthropicLimitedAccess, AnthropicSelfHostedEnvironmentAccess. Details are in the compliance-and-iam page (other worker).
• Data residency: inference_geo "us" (1.1x) or "global"; workspace allowed_inference_geos / default_inference_geo; workspace geo fixed to "us". Workspaces created before 2026-09-18 00:00 UTC may have been processed outside AWS.
• Billing: AWS Marketplace, CCUs; Start/Build/Scale tiers each with a monthly spend cap (HTTP 400 at cap until month rollover); self-set org/workspace spend limits. Private offers must be accepted before usage.
• Logging: CloudTrail (management events by default; inference/batch/file/skill/model/agent ops are Data events, extra cost). Response headers x-amzn-requestid (primary) + request-id.
• Managed Agents differences: 6-hour autonomous-session reauthentication (send any user event); no memory stores on self-hosted environments.

# 2.4 Google Cloud (Vertex AI / Agent Platform)

• POST https://{LOCATION}-aiplatform.googleapis.com/v1/projects/{PROJECT}/locations/{LOCATION}/publishers/anthropic/models/{MODEL}:rawPredict (regional, e.g. us-east5, europe-west1; Sonnet 4.6 and earlier only); global: https://aiplatform.googleapis.com/v1/projects/{PROJECT}/locations/global/...; multi-region: https://aiplatform.{us|eu}.rep.googleapis.com/v1/projects/{PROJECT}/locations/{us|eu}/.... Authorization: Bearer $(gcloud auth print-access-token). Body carries "anthropic_version": "vertex-2023-10-16", no model.
• Endpoint types: global (recommended, pay-as-you-go only), multi-region us/eu (+10 %), regional (+10 %, required for provisioned throughput). Pricing premium applies from Sonnet 4.5 onward.
• Ids: claude-fable-5-1, claude-fable-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5@20251101, claude-opus-4-1@20250805 (dep.), claude-opus-4@20250514 (dep.), claude-sonnet-5, claude-sonnet-4-6, claude-sonnet-4-5@20250929, claude-sonnet-4@20250514 (dep.), claude-haiku-4-5@20251001, claude-3-5-haiku@20241022 (dep.). Mythos Preview by invitation.
• Supported: Messages, prompt caching (auto + 5m + 1h), thinking, tool use incl. browser use and computer use toolsets, web search (basic version only), citations, structured outputs. Not supported: URL/Files sources, code execution, web fetch, advisor, Skills, MCP connector, programmatic tool calling, Batches/Models/Admin/Compliance/Usage endpoints, Managed Agents, server-side fallback. 1M context on Fable 5.1/5, Opus 5/4.8/4.7/4.6, Sonnet 5/4.6. Payload limit 30 MB.

# 2.5 Microsoft Foundry

• https://{resource}.services.ai.azure.com/anthropic/v1/*; model = deployment name (default = Claude API id). Two-level hierarchy: Foundry resource → deployments. Two hosting options chosen per deployment: Hosted on Azure (Global Standard or US Data Zone Standard = inference_geo: "us" equivalent, 1.1x) and Hosted on Anthropic (Global Standard only). Model version 1 = Hosted on Anthropic, version 2 = Hosted on Azure.
• Ids/deployments: claude-fable-5-1, claude-fable-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-5, claude-sonnet-4-6, claude-sonnet-4-5, claude-haiku-4-5 (* = also Hosted on Azure; all Hosted on Anthropic). Mythos Preview by invitation.
• Env: ANTHROPIC_FOUNDRY_API_KEY, ANTHROPIC_FOUNDRY_RESOURCE xor ANTHROPIC_FOUNDRY_BASE_URL (C# ignores the latter). SDKs: C#, Java, PHP, Python, TypeScript; not Go/Ruby.
• Not supported anywhere on Foundry: Admin API, advisor tool, Managed Agents, Compliance API, Models API, Message Batches, server-side fallback, computer/browser toolsets (beta computer-use versions remain). Additionally not supported when Hosted on Azure (400 by design): code execution, web search/fetch versions newer than web_search_20250305 / web_fetch_20250910, Agent Skills, programmatic tool calling, Files API.
• Response headers: request-id + apim-request-id; no anthropic-ratelimit-* headers. usage object identical across platforms. Billing: Azure Marketplace CCUs.

# 3. Regions and data residency

inference_geo is not applicable on Bedrock, Google Cloud and Foundry. Supported customer countries (api/supported-regions) is a jurisdiction list (≈175 countries), independent of cloud regions. ZDR/HIPAA: Anthropic programs apply to Claude API and (ZDR only, on request) Claude Platform on AWS; not to Bedrock/Google Cloud (cloud provider is processor); HIPAA program unavailable on Claude Platform on AWS and Foundry.

# 4. Pricing differences

• Bedrock and Google Cloud: partner price lists (aws.amazon.com/bedrock/pricing, cloud.google.com/vertex-ai/generative-ai/pricing); regional/multi-region +10 % over global for Sonnet 4.5 and later.
• Claude Platform on AWS and Foundry: same per-model/per-feature USD rates as the Claude API, converted to CCUs at $0.01/CCU after discounts; single CCU line item; no prepaid balance. Batch discount (50 %) applies where Batches exist (Claude API, Claude Platform on AWS). Fast mode premium pricing is Claude API only.
• Retired-on-API models still billed on partners: Opus 4.1, Sonnet 4, Haiku 3.5 (Bedrock + Google Cloud), Opus 4 (Google Cloud only).

# 5. Endpoint availability summary (count_tokens, batches, admin, agents, skills, files)

¹ Listed in the Features overview for all platforms; the exact REST path on Bedrock/Google Cloud/Foundry is not shown in Anthropic's docs (SDKs expose messages.count_tokens).

# 6. Feature × Platform matrix

Legend: Yes · No · Partial · Unknown (docs silent). Footnotes a–z after the table. Columns: API = Claude API, CPA = Claude Platform on AWS, BR = Claude in Amazon Bedrock (Mantle, Opus 4.7+), BRL = Bedrock legacy InvokeModel/Converse, GC = Google Cloud, MF = Microsoft Foundry. Beta header names in the JSON twin.

Footnotes: ᵃ basic web_search_20250305 only (no dynamic filtering). ᵇ Hosted on Azure: basic versions web_search_20250305 / web_fetch_20250910 only; Hosted on Anthropic: all versions. ᶜ Hosted on Anthropic deployments only; Hosted on Azure returns 400. ᵈ computer_toolset_20260801 unavailable; earlier beta versions computer_20251124 (computer-use-2025-11-24) and computer_20250124 (computer-use-2025-01-24) remain. ᵉ Contradiction: the Bedrock (Mantle) page lists structured outputs under "not supported", the Features overview lists Bedrock with a footnote limited to Opus 4.6 / Sonnet 4.6 / Sonnet 4.5 / Opus 4.5 / Haiku 4.5 — i.e. the legacy-integration models. ᶠ Header accepted on any model without error, ignored where unsupported (Haiku 4.5 never supports it). ᵍ 6-hour autonomous-session reauthentication; no memory stores on self-hosted environments. ʰ 1M on Fable 5.1/5, Opus 5/4.8/4.7/4.6, Sonnet 5/4.6 (Bedrock Mantle: not Opus 4.6 which is legacy-path); Haiku 4.5 and Sonnet 4.5 = 200k. ⁱ Token bound to caller identity (no workspaces) on Bedrock/Google Cloud; allowed_fallback_models not exposed on partners/Foundry; fallback_credit field may be absent while rollout continues. ʲ Features overview marks server-side fallback as Claude API beta only; CPA page does not list it as unsupported — treated as No. ᵏ Thinking/mid-conversation pages say "on Amazon Bedrock, Google Cloud, and Microsoft Foundry, pass the beta value as described in Beta headers", but the Beta headers page in this snapshot only documents the anthropic-beta header; no body-level anthropic_beta array is documented anywhere in the corpus (grep anthropic_beta = 0 hits). Bedrock comparison table states the header is "not supported".

# 7. Model × Platform id table

Status column = Claude API lifecycle (about-claude/model-deprecations); partner platforms set their own dates (Bedrock/Google), Foundry and Claude Platform on AWS follow Anthropic's.

Id-format rules (model-ids-and-versions): 4.6+ dateless claude-{name}-{major}[-{minor}] ↔ Bedrock anthropic.claude-… (Opus 4.6 is the last with -v1) ↔ Google identical; pre-4.6 dated claude-…-{YYYYMMDD} ↔ Bedrock anthropic.claude-…-{YYYYMMDD}-v1:0 ↔ Google claude-…@{YYYYMMDD}. Dateless ids are pinned snapshots, not evergreen aliases. Claude Code /claude-api migrate rewrites ids per platform.

# 8. Notable gaps and contradictions

1. Structured outputs on Bedrock: platform page (Mantle) says not supported; Features overview says supported with a model footnote covering only legacy-path models. Recorded as No (Mantle) / Yes (legacy, 5 models).
2. Beta mechanics on partner platforms: several feature pages point to "Beta headers" for Bedrock/Google Cloud/Foundry, but that page has no partner section and no anthropic_beta body field exists in the corpus; the CPA comparison table says the header is not supported on Bedrock. Yet Bedrock/Google Cloud are listed as (beta) for compaction, context editing, fallback credit, computer-use beta versions and accept interleaved-thinking-2025-05-14. How those betas are switched on there is undocumented → Unknown.
3. count_tokens paths on Bedrock, Google Cloud and Foundry are not shown (only "available" in the overview) → endpoint records flagged DOCUMENTATION_INCOMPLETE.
4. Google Cloud streaming REST method (:streamRawPredict) is not in the Anthropic snapshot; only the SDK is documented as streaming.
5. Foundry supports several betas (MCP connector, Skills, Files) but the Beta-header classification note names only Claude API and Claude Platform on AWS.
6. Rate-limit response headers are documented only for the Claude API (present) and Foundry (absent).
7. Model availability nuance: Claude Platform on AWS lists no Mythos models; Bedrock Mantle omits Opus 4.6 / Sonnet 4.6 / 4.5 / Opus 4.5 (legacy path only); Foundry hosts Fable, Opus 4.7/4.6/4.5, Sonnet 4.6/4.5 on Anthropic infrastructure only.
8. Retired-on-API models (Opus 4.1, Opus 4, Sonnet 4, Haiku 3.5, Sonnet 3.7) keep separate partner lifecycles; Bedrock already lists Sonnet 3.7 and Opus 4 as retired while Google Cloud still lists Opus 4.

# 9. Examples (all UNVERIFIED, Haiku 4.5, max_tokens: 8, "Reply with OK.")

examples/anthropic/cloud/{bedrock,vertex,foundry}_messages.{py,ts,sh} — records in tmp/platform-anthropic/examples-cloud.json. Bedrock scripts use AnthropicBedrockMantle / --aws-sigv4 bedrock-mantle with the legacy InvokeModel path in comments; Vertex uses AnthropicVertex / gcloud auth print-access-token + :rawPredict; Foundry uses AnthropicFoundry / api-key with Entra ID alternative.