Anthropic’s Commercial Terms of Service are the agreement under which businesses use Anthropic API keys and “any other Anthropic offerings that references these Terms”, together with related tools, documentation and services.[1] They are short, and much of their weight sits in documents they incorporate by reference: the Usage Policy, the Supported Regions Policy, the Service Specific Terms and the Data Processing Addendum (DPA).[1] A compliance review for a Claude deployment therefore reads five documents with different effective dates, each of which Anthropic updates on its own. Plan, seat and token pricing is covered in Anthropic Claude licensing, and the summary of the whole contract in Anthropic licensing.
Editions
Commercial or consumer terms
The Commercial Terms state that “Services under these Terms are not for consumer use” and that consumer offerings such as Claude.ai are governed by the Consumer Terms of Service instead.[1] The contracting entity depends on where the customer resides: Anthropic Ireland, Limited for the EEA, Switzerland and the UK, and Anthropic, PBC elsewhere. The same split decides arbitration, with a sole arbitrator in Dublin under the UNCITRAL Rules or in San Francisco under JAMS rules.[1] The Terms take effect on the earlier of first electronic consent and first access, and only a person with authority to bind the organization may accept them.[1]
Document map
| Document | Effective | What it adds |
|---|---|---|
| Commercial Terms of Service | 2025-06-17 | Use right, content ownership, restrictions, fees, termination, indemnity |
| Usage Policy | 2025-09-15 | Prohibited uses; High-Risk Use Case Requirements; guidelines for chatbots, minors, agents and MCP servers |
| Service Specific Terms | 2026-08-31 | Claude for Work, beta, fine-tuning, marketplace, Development Partner Program, Covered Models, trusted-access programs |
| Data Processing Addendum | 2025-02-24 | Controller and processor roles, subprocessors, audits, deletion, transfer clauses |
| Supported Regions Policy | Undated | Countries where commercial API access and Claude.ai are offered |
Sources: the five documents.[1][2][3][4][5] Where the Service Specific Terms conflict with the Commercial Terms, the Service Specific Terms “control for the relevant Services”.[3]
Metrics
The Commercial Terms set no licence metric of their own. The customer pays for fees incurred by its account “at the rates specified on the Model Pricing Page, unless otherwise agreed”, and Anthropic may require prepayment in credits under its Supplemental Credits Terms.[1] For use through cloud marketplaces, fees are paid to the Marketplace Provider at Anthropic’s then-current rates unless otherwise agreed.[3] Rate changes take effect at the earlier of 30 days after posting or Notice to the customer, under both documents.[1][3]
Counting / floors
Who may use the Services
The use right covers the customer and also extends to powering “products and services Customer makes available to its own customers and end users”, whom the Terms call Users.[1] The customer is responsible for all activity under its account.[1] The Users of a customer’s product are therefore within the licence, but they bind the customer: the Usage Policy and other incorporated policies apply to “Customer and its Users”.[1]
For Claude for Work (the Team and Enterprise plans), the customer must tell its Users that they are using an administered service that gives the customer access to and control over their data, and that use is subject to Anthropic’s Privacy Policy and Usage Policy. Obtaining User consents and giving legally required notices is the customer’s job.[3]
Use restrictions
The customer may not use the Services to build a competing product or service, “including to train competing AI models or resell the Services except as expressly approved by Anthropic”, may not reverse engineer or duplicate them, and may not support a third party in doing so.[1] Catalog rule: Commercial use restrictions. Training a model through Anthropic’s own Fine-Tuning Services is carved out: it does not breach the training prohibition “provided that the training conducted is solely for Customer’s use of the Fine-Tuning Services”.[3]
Outputs
The customer must judge whether Outputs suit its use case, including where human review is appropriate, and must notify its Users that factual assertions in Outputs should not be relied on without checking.[1] The customer retains its Inputs and owns its Outputs, and Anthropic may not train models on Customer Content.[1] Catalog rule: No training on Customer Content.
Usage Policy
The Usage Policy, formerly the Acceptable Use Policy, has three layers.[2]
Universal Usage Standards apply to all users and use cases. They are grouped under fourteen headings, from “Do Not Violate Applicable Laws or Engage in Illegal Activity” to “Do Not Generate Sexually Explicit Content”, and include prohibitions on compromising computer or network systems, developing weapons, undermining democratic processes and certain criminal justice, surveillance and law enforcement uses.[2] Anthropic may contract with certain government customers for use restrictions tailored to their public mission if it judges the restrictions and safeguards adequate.[2]
High-Risk Use Case Requirements apply to legal, healthcare, insurance, finance, employment and housing, academic testing and admissions, and media or professional journalistic content. For these, two measures are mandatory:[2]
- Human-in-the-loop. Where the Services give advice or recommendations, or support subjective decisions, directly affecting individuals or consumers, “a qualified professional in that field must review the content or decision prior to dissemination or finalization”.
- Disclosure. Where outputs are presented directly to individuals or consumers, the customer must disclose that AI is used, at a minimum at the beginning of each session.
Wellness advice on matters such as sleep, stress, nutrition or exercise is outside the healthcare category.[2]
Additional Use Case Guidelines apply whether or not a use case is high-risk. All consumer-facing chatbots, “including any external-facing or interactive AI agent”, must disclose at the start of each session that users are interacting with AI. Products serving minors must follow further guidelines in Anthropic’s Help Center, agentic use must still comply with the Usage Policy, and MCP servers in Anthropic’s Connector Directory must follow its Directory Policy.[2]
Enforcement
Anthropic’s Safeguards Team runs detection and monitoring, and for violations Anthropic “may throttle, suspend, or terminate your access” and may block or modify outputs.[2] Under the Commercial Terms, Anthropic may suspend access where a customer or User violates the compliance, policy or use-restriction sections, and the customer must cooperate with reasonable information requests, including identity verification.[1] The customer indemnifies Anthropic against third-party claims related to use “in violation of the Usage Policy, the Service Specific Terms, or Section D.4”.[1] Anthropic’s own IP indemnity covers “paid use” only and excludes customer modifications, combinations with non-Anthropic technology and customer Inputs.[1]
Cloud / BYOL
The Service Specific Terms name two Cloud Platforms that host the Services: Amazon Bedrock, run by AWS, and Vertex AI, run by Google. Anthropic’s enabling technology does not give it access to customer instances on those platforms, subject to the Covered Models terms.[3] Customers must also follow the Marketplace Provider’s own policies and agreements, and Anthropic is not responsible for the platforms.[3]
The Supported Regions Policy lists the countries where commercial API access and Claude.ai are offered. To the extent the law permits, Anthropic may decline to serve entities whose majority direct or indirect ownership is attributable to nations not on the list.[5] Ownership, not only location, therefore matters for eligibility.
Programs
Beta, Development Partner and Covered Models
- Beta Services are provided “as-is” on a temporary basis, carry no indemnity, and limit Anthropic’s liability to the lesser of $1,000 and fees paid in the previous 12 months.[3]
- Development Partner Program. A customer that enables Development Partner Mode allows Anthropic to use the data it submits, “including to train models”.[3] This is the one setting that reverses the no-training commitment in the Commercial Terms, so it belongs in any data review.
- Covered Models. For models Anthropic designates as Covered Models, Anthropic may retain and safety-review Inputs, Outputs and usage data, and this “supersedes any modified retention commitments (e.g., ZDR)”.[3]
- Trusted Access and Verification Programs may give access to models or capabilities only to participants, on separate program terms.[3]
Data Processing Addendum
Data submitted through the Services is processed under the DPA, which is incorporated into the Commercial Terms.[1] For Customer Personal Data the customer is controller and Anthropic processor.[4] The customer gives general authorization for the listed Subprocessors, and Anthropic remains liable for them within the agreed liability cap.[4] Audit rights are met through annual external audits: on written request Anthropic provides audit reports or certificates such as a SOC 2 report.[4] Within 30 days of termination Anthropic returns Customer Data if asked within that period and deletes it, except where law requires retention, to resolve a dispute, or “to combat harmful use of the Services”.[4] International transfers use the EU Standard Contractual Clauses, Modules Two and Three, and the UK Addendum.[4]
Changes and termination
Anthropic may update the Commercial Terms with effect 30 days after posting or Notice, or immediately for changes required by law, and “Changes will not apply retroactively”.[1] Either party may terminate for convenience, Anthropic with 30 days’ prior Notice, and confidentiality, fees, indemnification and several other sections survive termination.[1]
Out of scope
- The Consumer Terms of Service and individual Pro and Max plans.
- Negotiated enterprise agreements, which can vary the published terms.
- The full list of supported countries and of Subprocessors.
- Terms of Amazon Bedrock, Vertex AI and other Marketplace Providers.