General-purpose · General-purpose assistants · High confidence
Anthropic, PBC · First-party Claude API and Claude Platform (api.anthropic.com). ZDR and HIPAA-ready are optional and feature-limited. Amazon Bedrock, Google Vertex AI and Microsoft Foundry need separate profiles/contracts.
Developer API for Anthropic models. Zero data retention is available on eligible endpoints only. The wrapping application owns retrieval, logging, identity and human gates.
Opens a private software assessment of Claude API for one task, with this profile’s questions to hand.
| Deployment | SaaS, Private cloud |
|---|---|
| Hosting regions | United States, Worldwide or dynamically routed |
| Australian region | No |
| Model providers | Anthropic |
| Assurance | SOC 2 Type II, ISO 27001:2022, ISO/IEC 42001:2023, CSA STAR, NIST 800-171 |
| Training on customer content | No |
| Retention summary | Standard API deletes inputs/outputs within 30 days. ZDR available on request for eligible Messages/Token Counting features and Claude Code; many features excluded. Covered Models require 30-day retention. Activity Feed 6 years. |
| Integrations | Amazon Bedrock (separate contract), Google Vertex AI (separate contract), Microsoft Foundry (separate contract) |
| Pricing summary | Published token pricing by model. Optional US-only inference multiplier. Batch discount advertised. Console/API credits or invoiced spend depending on account. |
| Legal content coverage | None native. Any legal grounding is supplied by the wrapping application. |
| Australian presence | Yes — API access from Australia is supported. Workspace geo is US-only. Named AU customers may use Platform, Team or Enterprise; not confirmed as API-only. |
Define the configuration, people, information and tasks being assessed.
Public claim only
Direct Claude API (api.anthropic.com) and Claude Platform, including optional Claude Platform on AWS. Bedrock, Vertex and Foundry are separate contracts with the cloud provider as processor. ZDR and HIPAA-ready are optional arrangements with feature eligibility tables. Console playground is outside ZDR.
Ask: Confirm we are contracting first-party API versus Bedrock/Vertex/Foundry, and list enabled endpoints, tools, Files API, Batch, Managed Agents and MCP.
Public claim only
API SaaS. Workspace geo is currently US-only and cannot be changed after workspace creation. Inference geo is global (default) or us, on Claude 4.6+ models, at 1.1x price for US-only. Private-cloud style options are Bedrock/Vertex/Foundry under other agreements.
Ask: Record workspace geo, default_inference_geo and allowed_inference_geos in the order form.
Public claim only
Messages, Token Counting, Files, Batch, Skills, Managed Agents, code execution, web search/fetch, computer use and MCP connectors. Several of these are excluded from ZDR (Console, Managed Agents, Claude for Excel, consumer products, some Covered Models, CORS, third-party integrations).
Ask: Which API features will be enabled, and which of them fall outside ZDR or HIPAA-ready?
Public claim only
Australia is supported for API access. There is no Australian workspace geo. Global routing may process inference in selected countries including Australia; storage remains US.
Ask: Is US storage plus optional AU inference routing acceptable under our APP 8 and client guidelines?
Record the practical work needed to introduce and maintain the software.
Public claim only
Published per-million-token rates by model (for example Opus 5 input USD 5 / output USD 25 per MTok at review date). US-only inference is 1.1x on Claude 4.6+. Batch processing advertised at 50% off. Commercial Terms: Anthropic may update published rates on 30 days' notice.
Ask: Lock model prices, the 1.1x residency multiplier and any spend caps in the order form.
Not confirmed
The API has no Word add-in of itself; the customer application must implement identity, logging and save-back. Console organisations can use SSO/JIT/SCIM (SCIM on Enterprise/Console parent orgs).
Ask: Who owns the wrapping application (internal build, Harvey, Legora, other) and does that wrapper's contract flow down these API terms?
Public claim only
Developer platform support; security via Responsible Disclosure / HackerOne. No public API uptime SLA was inspected in this review. DPA audit rights exist.
Ask: What is the API availability SLA, status page commitment and security severity response time?
Compare complete work at the required quality, including checking and correction.
Public claim only
Model documentation forms and training-data summaries are listed on the Trust Center. Commercial Terms still require the customer to evaluate outputs. Legal accuracy depends on the wrapping application.
Ask: What groundedness tests will the wrapping application run on Australian authorities?
Not confirmed
The API is a model endpoint. Task fit is determined by the customer application (prompts, retrieval, citations, human gates). Using the raw API for advice or court documents without those controls is a firm design choice, not a product guarantee.
Ask: Document the wrapping application's retrieval sources, citation checks and refusal behaviour for our tasks.
Compare the actual configuration with firm, client and jurisdiction requirements, supported by evidence.
Not confirmed
If the firm exposes Claude via a client portal, unqualified-practice and disclosure rules fall on the firm. Commercial Terms require notifying users that outputs may be false.
Ask: Will any end-client application call this API, and what practitioner review is mandatory before the client sees output?
Public claim only
Anthropic. Covered Models (including named Fable/Mythos versions) require 30-day retention and are not ZDR-eligible unless expressly authorised. Customer can pin models per request.
Ask: Which model IDs are approved, and will Covered Models be disabled if we require ZDR?
Public claim only
Published list (GCP, AWS, Azure, Cloudflare, and feature services). DPA flow-down and 15-day objection. Web search (Brave, TurboPuffer) applies if that tool is enabled.
Ask: Map each enabled API tool to a subprocessor and disable unused tools.
Public claim only
Commercial Terms: Anthropic may not train on Customer Content from the Services. Commercial privacy article: API inputs/outputs not used to train by default. Feedback/bug reports remain an exception if submitted.
Ask: Confirm no-training in the signed API agreement, including embeddings and evaluation datasets.
Public claim only
Console thumbs-up/down and explicit bug reports can retain conversations up to 5 years and allow training. Avoid sending matter content through those channels.
Ask: Disable or administratively forbid feedback from production API keys.
Public claim only
Standard API: inputs/outputs deleted within 30 days except Files API, agreed otherwise, ZDR, Covered Models, Usage Policy flags or law. ZDR: Anthropic does not store prompts/responses at rest after the response is returned, per organisation, on eligible Messages/Token Counting features and Claude Code with commercial keys; must be enabled by sales. Console, Managed Agents, consumer products, Team/Enterprise UIs, Claude for Excel, and several tools are outside ZDR. Activity Feed retains 6 years.
Ask: Enable ZDR in writing for our organisation, list excluded features we will not call, and state Covered Model handling.
Public claim only
DPA 30-day return/delete on termination. ZDR organisations have little stored content to delete except excluded features, flags and legal holds. Files API and Managed Agents persist until deleted by the customer.
Ask: How are Files, batches, agent transcripts and abuse-monitoring stores deleted on instruction, with what certificate?
Public claim only
Access Transparency can log Anthropic personnel views of covered Messages API and Claude Code content to the Compliance API (eligibility required; does not cover claude.ai apps, consumer, Files/Batch/non-ZDR features, or Bedrock/Vertex). Automated safety processing does not emit those events.
Ask: Enable Access Transparency for our API organisation and confirm support staff cannot read prompts without a logged reason code.
Public claim only
API keys and Console roles (Admin, Developer, Billing, etc.). SSO/JIT/SCIM available for Console organisations on documented tiers. Key handling is the customer's responsibility.
Ask: How are production keys scoped, rotated, stored in our vault, and revoked on leaver?
Public claim only
DPA AES-256 at rest, TLS 1.2+. CMEK available for Claude Platform (per workspace) in US regions; some Console features disable when CMEK is on. ZDR is compatible with CMEK.
Ask: Will CMEK be attached to a new US workspace before any prompts are sent?
Public claim only
Trust Center: Claude via Anthropic API in scope for SOC 2 Type II, ISO 27001, ISO 42001, CSA STAR, HIPAA, NIST 800-171. Reports on request. Not inspected here.
Ask: Provide current reports scoped to the first-party API, including AI features we will call.
Public claim only
Commercial Terms + DPA. Customer owns Outputs. 12-month fee cap. California / SF arbitration for Australian customers. Rate card may change on 30 days' notice.
Ask: Execute DPA, confirm no-training and ZDR on the order form, and negotiate cap/forum if required.
Public claim only
US workspace storage. Inference global or US. SCCs in DPA. No Australian region. Cloud-provider routes have their own regions.
Ask: Document inference_geo for production traffic and complete an APP 8 assessment for US storage.
Public claim only
DPA 48-hour Security Breach notice. Status page. Responsible disclosure policy. Evaluation-environment incidents have been published separately by Anthropic.
Ask: Confirm 48-hour notice and NDB cooperation for API tenants, plus status-page SLA.
Public claim only
Terms and prices updatable on 30 days' notice. Model deprecation commitments exist. Feature eligibility for ZDR/HIPAA can change. Covered Model list can expand.
Ask: What notice before a model snapshot is retired or a tool loses ZDR eligibility?
Public claim only
DPA 30-day export/delete. API customers can export via their own logs if they captured them. Activity Feed is a vendor-side record with a long retention.
Ask: Will we receive deletion of Files, agents, activity feed and abuse stores, with a certificate, within 30 days?
Public claim only
API tools (web search, MCP, computer use, code execution) send data to subprocessors or customer-configured MCP servers. They are often outside ZDR. The wrapping app must not enable them by default for privileged matter.
Ask: Which tools are allowed on production keys, and are MCP servers only our own allow-listed endpoints?
Public claim only
Web search is an API tool using Brave Search / TurboPuffer. Not on by default unless the application enables it. Outside some ZDR/HIPAA feature tables.
Ask: Keep web_search disabled on production keys unless a named non-confidential task requires it.
Not confirmed
Managed Agents persist transcripts. Computer use, bash and MCP can act outside the model. Human confirmation is an application design choice. Anthropic can suspend for Usage Policy or attack.
Ask: Which tools require an application-level human confirm, and how do we revoke keys and halt agents immediately?
Public claim only
Commercial Terms place evaluation of outputs on the customer. Compliance API Activity Feed (6 years) and Access Transparency support supervision if enabled. The API will not itself refuse to answer a legal question.
Ask: How will our wrapping application log prompts/outputs into our matter file and force lawyer review before use?
Official pages reviewed for this profile, last reviewed 2026-09-20.
Add this product to your toolset for a particular task, or use it with a practice-area checklist. Add to your toolset Practice areas Best AI general-purpose assistants tools Australia
