Anthropic Claude API Terms of Service Limits
If you're searching for "Anthropic Claude API terms of service limits," you're probably trying to answer one of two questions: how many requests can I actually make, or am I allowed to build and sell a product on top of this. Both questions have different answers, because Anthropic's terms combine technical rate limits (tied to your usage tier) with policy restrictions (what you're allowed to do with the model's output and how you can commercialize access to it).
This article separates the two. First we cover the hard technical limits — requests per minute, tokens per minute, context window — and then the contractual limits from Anthropic's Terms of Service and Acceptable Use Policy that govern what kind of product you can legally build.
Technical Limits vs. Policy Limits
It helps to think of Claude API restrictions in two buckets:
- Technical limits: rate limits (RPM, TPM), context window size, max output tokens, concurrent request caps. These are enforced automatically by the API and return HTTP 429 or 400 errors when exceeded.
- Policy limits: restrictions defined in the Terms of Service and Acceptable Use Policy — what content you can generate, who you can serve, and how you can resell or redistribute access.
Developers usually hit technical limits first (a spike in traffic returns 429s), but policy limits matter more when you're deciding whether your product idea is even allowed.
Rate Limits by Usage Tier
Anthropic assigns accounts to usage tiers based on spend history and account age. Each tier raises your requests-per-minute (RPM), tokens-per-minute (TPM), and tokens-per-day (TPD) ceilings. New accounts start on the lowest tier and move up automatically as billing history accumulates — there's no manual "request an increase" form for most tiers, though enterprise agreements can negotiate custom limits.
Typical patterns you'll run into:
- Low tiers cap you at a handful of requests per minute — fine for prototyping, not for production traffic.
- Token-per-minute limits usually bind before request-count limits once you're sending longer prompts or streaming responses.
- Limits apply per organization, not per API key, so spreading traffic across multiple keys inside the same account doesn't help.
Because these limits scale with billing history rather than a form submission, teams that need predictable throughput from day one often route traffic through a managed layer instead of waiting out the tier progression. This is one of the reasons tools like SubToAPI exist — they sit on top of your existing Claude access and give you a stable sub_live_... API key, usage metadata, and streaming without you having to manage tier math yourself. See the quickstart for how that mapping works.
Context Window and Output Limits
Separate from rate limits, each Claude model has a fixed context window (input + output tokens combined) and a maximum output token limit per response. These aren't negotiable by tier — they're a property of the model itself. If your terms-of-service question is really about "why did my request get rejected even though I'm under my rate limit," the answer is usually a context window overflow, not a ToS violation. Check the messages documentation for current per-model limits before assuming it's a policy issue.
Acceptable Use Policy Restrictions
The Acceptable Use Policy is the part of the terms most developers skip and shouldn't. It defines categories of use that are restricted or outright prohibited regardless of your tier or spend level, including:
- Generating content related to weapons development, especially CBRN (chemical, biological, radiological, nuclear) material.
- Building surveillance tools that violate privacy rights or enable mass tracking without consent.
- Use in high-risk domains — medical diagnosis, legal advice, financial decisioning — without appropriate human review and disclaimers.
- Generating deceptive content: deepfakes, impersonation, coordinated disinformation.
- Child safety violations, which Anthropic treats as zero-tolerance.
If your product touches any "high-risk" category (healthcare, legal, financial, political), the terms generally require that a human review AI-generated output before it reaches an end user, and that you disclose AI involvement. This isn't a rate limit, but violating it can get your account suspended just as fast as hitting a hard cap — so it belongs in the same conversation.
Commercial and Resale Terms
A common question: can I build a SaaS on top of Claude and charge customers for it? Yes — this is explicitly a supported use case, and Anthropic's commercial terms are built around exactly this. What the terms restrict is different from what people assume:
- You generally can't present your product as being Anthropic's own first-party offering, or imply an endorsement/partnership that doesn't exist.
- You're responsible for how your end users use the model through your product — the Acceptable Use Policy flows down to them.
- Reselling raw, unrestricted API access with no added value or safeguards is treated differently from building an actual product (features, UI, authentication, usage controls) on top of the API.
This last point is why most teams add a layer between their Anthropic account and their end users rather than exposing a bare API key to customers. A tool like SubToAPI is designed for exactly that pattern: it wraps your Claude access in application-scoped API keys, lets you issue separate keys per customer or per environment, and gives you usage metadata per key — so you're not just forwarding a single root credential to everyone who uses your product. Pricing starts at the Solo plan (€9) for individual builders, with Team (€19/seat) and Scale (€49/seat) plans for products with multiple users; see pricing.
Data Retention and Training Use
The terms also define how your prompts and outputs are handled. For API usage (as opposed to the consumer chat product), Anthropic's standard commercial terms state that your data isn't used to train their models by default, and retention periods for abuse-monitoring purposes are limited and disclosed in the terms. If your product handles sensitive customer data, read the current Data Processing terms directly from Anthropic rather than relying on second-hand summaries — this is the part of the ToS that changes most often as regulatory requirements evolve.
Staying Inside the Limits in Practice
The most common real-world failure mode isn't a policy violation — it's an unhandled 429 during a traffic spike. A minimal retry pattern:
async function callWithBackoff(fn, retries = 5) {
for (let attempt = 0; attempt < retries; attempt++) {
try {
return await fn();
} catch (err) {
if (err.status === 429 && attempt < retries - 1) {
const wait = 2 ** attempt * 500;
await new Promise(r => setTimeout(r, wait));
continue;
}
throw err;
}
}
}If you're routing requests through SubToAPI, the same pattern applies against https://api.subtoapi.app/v1/messages — rate limit responses carry the same structure, and the streaming docs cover backoff behavior for long-running connections.
Questions
Does exceeding my rate limit count as a Terms of Service violation? No. Rate limits are technical, enforced automatically, and resolve on their own — they're not a policy breach. ToS violations relate to prohibited content or misuse categories in the Acceptable Use Policy.
Can I legally resell Claude API access to my own customers? Yes, building and charging for a product on top of the API is supported, as long as you're adding real functionality and your end users remain bound by the Acceptable Use Policy through your own terms.
Will my API usage be used to train Anthropic's models? Standard commercial API terms exclude your prompts and outputs from model training by default — confirm current wording in Anthropic's data processing terms, since this can vary by agreement type.