Two model providers changed their usage policies in the same week. Anthropic's updated Usage Policy took effect on November 12, 2026, and OpenAI's takes effect on November 20, 2026. Most coverage leads with gateways and with abuse toward Claude, but a startup has a narrower question: does any of this change what my product may do? This guide compares the two updates, shows when OpenAI now asks for approval, explains what both say about training on model outputs, lists the marketing features Anthropic's new deceptive-activity section reaches and ends with a self-check you can copy. It is useful without a Soloop account. It is not legal advice and not a statement from OpenAI or Anthropic.
The short answer
If your product uses OpenAI or Claude models to power its own features, the November updates do not ask you to apply for anything. OpenAI's approval rule is for services whose job is giving other people direct access to its models, such as model routers and AI gateways, and its policy says the rule does not apply just because an app integrates its models.
Three things deserve a check this week. First, do not use either provider's outputs to train or tune a model that replaces theirs without permission. Second, if your product posts, reviews or publishes for customers, read Anthropic's new rules on fake personas, astroturfing and seeding AI answers. Third, if your product makes recommendations about money, jobs, housing, health, education or legal matters, keep a qualified person reviewing them and tell people AI was used.
What changed, side by side
Anthropic announced its update on October 8, 2026, and OpenAI announced its update on October 21, 2026. Anthropic says most of its changes clarify existing rules. OpenAI's changes add new approval requirements and spell out the training rules in more detail.
| Dimension | OpenAI (effective November 20, 2026) | Anthropic (effective November 12, 2026) |
|---|---|---|
| Giving others access to the model | A model-access intermediary, or any service that lets third parties directly access OpenAI models, needs OpenAI's prior review and approval. Integrating the models into your own app does not trigger this on its own. | Reselling, proxying or otherwise providing access to Claude through unauthorized means is prohibited, including routing requests through consumer subscriptions or misrepresenting the product or client being used. |
| Training on outputs | You may not use OpenAI models or outputs to prepare, select or supply training material or signals for models intended as alternatives to or substitutes for OpenAI's. Research, evaluation, benchmarking and product development remain allowed. | You may not use outputs to train an AI model, such as through model scraping or distillation, without Anthropic's prior authorization. |
| Deceptive and artificial activity | Deceit, fraud, scams, spam and impersonation remain prohibited, as does using someone's likeness or voice without consent in ways that could confuse authenticity. | A new section covers fake personas and accounts, hidden sponsorship, networks of ostensibly independent outlets, seeding the sources AI systems answer from, and building or selling tools for these. |
| High-stakes decisions | No automation of high-stakes actions without human review in areas such as education, housing, employment, credit, insurance, legal and medical. Licensed advice needs a licensed professional, and the AI may not claim credentials. | The requirements are unchanged but rewritten: a qualified person must meaningfully review a high-risk recommendation and be able to change it, and the person affected must be told AI was used. |
| Other changes | Clarified protections for people under 18; national security and law enforcement uses need OpenAI's approval; OpenAI is introducing SafetyIDs with customers to tell end users apart in investigations. | A ban on sustained, needless abuse toward Claude; controls for hardware that takes physical actions; clearer Supported Regions enforcement. |
Effective November 20, 2026; changelog dated October 21, 2026 describes the training, teen, high-stakes, national security and model-access intermediary changes.
OpenAI Usage PoliciesServices that provide direct access to OpenAI models, such as model routers or AI gateways, need review and approval; the policies apply through hosting and reseller partners too; SafetyIDs are being introduced with customers.
OpenAI: Broad access through consistent standards (Oct 21, 2026)Effective November 12; new section on deceptive campaigns or artificial activity, political or commercial; high-risk requirements unchanged but rewritten; abuse toward models; physical actions; Supported Regions.
Anthropic: 2026 Usage Policy update (Oct 8, 2026)Is your product an app or a gateway?
OpenAI does not publish a list of approved gateways or a test beyond its policy sentence. The table applies that sentence to common startup setups. Rows marked inference are this guide's reading, not OpenAI's.
| Dimension | Approval needed? | Why, and what to check |
|---|---|---|
| A SaaS feature that calls the model on the user's behalf (summaries, drafts, classification) | No. | The policy says the rule does not apply solely because an application integrates OpenAI models to power its features. |
| A chat or agent product where users type anything, built around your own workflow | Probably not (inference). | It is still your app's feature. Check that you are not mainly selling unrestricted model access under your brand. |
| An OpenAI-compatible API, model router or gateway that other developers call | Yes. | This is the case the policy names. OpenAI has not published how to apply; contact OpenAI before November 20 and do not assume continued access. |
| A bring-your-own-key tool where customers paste their own OpenAI key | Unknown. | The policy does not address it. Each customer's use runs under their own account, but a tool that simply forwards requests can look like an access service. Ask OpenAI if that is your product's core. |
| An agency or freelancer building on the client's own OpenAI account | No, in most cases (inference). | The client is the customer and must follow the policies; you are building their app, not offering access to others. |
| A product that relays Claude through a Pro or Max subscription instead of the API | Prohibited by Anthropic. | Anthropic's policy names services that route requests through consumer subscriptions. Use the API or an authorized provider. |
"If you operate a model-access intermediary platform or other service that enables third parties to directly access OpenAI models, you must obtain our prior review and approval. This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality."
OpenAI Usage PoliciesProhibits reselling, proxying or otherwise providing access to Claude through unauthorized means, including services that route requests through consumer subscriptions.
Anthropic Usage PolicyCoverage notes that OpenAI published no approved-gateway list and no application process for existing intermediaries before November 20.
eWeek (Oct 26, 2026)Training on model outputs
A common way to cut costs is to log a large model's answers and use them to tune a smaller one. Both updated policies restrict this, in different ways. OpenAI's rule depends on purpose: it bans using its models or outputs to build training material or signals for models intended as alternatives to or substitutes for OpenAI's models or services. Anthropic's rule is broader in wording: no training an AI model on outputs without its prior authorization.
A smaller model tuned on GPT answers so your product can stop calling GPT is, by plain reading, intended as a substitute for that model. Treat that as needing permission unless your agreement says otherwise (inference).
- List every place your product stores model outputs, including logs, evaluation sets and user-rated answers.
- For each one, write down whether anyone uses it to create training data, labels, rankings, rewards or filtered datasets, including a contractor or a vendor.
- Remove or separate any dataset meant to train a replacement model, or get written permission from the provider first.
- Keep evaluation, benchmarking and prompt testing, which OpenAI's policy still allows within its limits.
- Check your agreement with each provider for any permitted exception, and do not rely on a contractor doing the training for you; OpenAI's rule applies to work done on your behalf.
Covers creating training data, tasks, rubrics or reinforcement-learning environments; feedback such as labels, scores, rankings, preferences, rewards or corrections; and selecting, filtering or validating data. Applies whether you do the work yourself, for another party or through someone on your behalf, even if you do not train the model, and to outputs from product use later supplied for training.
OpenAI Usage PoliciesProhibits using outputs to train an AI model, such as model scraping or model distillation, without prior authorization from Anthropic.
Anthropic Usage PolicyMarketing features that need a second look
Anthropic moved its scattered rules on fake accounts and disinformation into one section that applies to commercial activity as well as politics. It covers using Claude to run deceptive activity and building the tools for it, and marketing or selling such tools. If you build an AI marketing, social or SEO product, check each feature against this list.
- Multiple accounts or personas that post as if they were different, independent people.
- Generated reviews, testimonials or comments that pretend to come from customers.
- Bots that say they are human or answer "are you a bot?" untruthfully.
- Sponsored content that hides who paid for it.
- Networks of sites or accounts that look unrelated but publish from one source.
- Pages created to be cited by search engines or AI answers while hiding their origin, for example sites posing as independent sources that confirm the same claims.
- Content that copies a real person's voice or face, or impersonates a company or outlet.
Prohibits fake personas, accounts, outlets or organizations and impersonation to mislead about origin, how widely a view is held or whether people are talking to a human (sockpuppets, astroturfing, fake reviews, bots that claim to be human); concealed sponsorship; networks of ostensibly independent outlets; manipulating the sources search engines or AI systems draw answers from; building tools for astroturfing, sockpuppet creation, coordinated inauthentic behavior or scaled impersonation; and marketing or selling them.
Anthropic Usage PolicyProhibits deceit, fraud, scams, spam or impersonation, and using someone's likeness, including their photorealistic image or voice, without consent in ways that could confuse authenticity.
OpenAI Usage PoliciesHigh-stakes recommendations: a person and a disclosure
If your product helps decide something that materially affects a person, both providers expect a person to stay accountable. Anthropic's rewritten section now lists which recommendations are covered, which makes the check easier.
| Dimension | Examples covered | What to build in |
|---|---|---|
| Hiring and work | Screening, ranking or rejecting candidates; setting pay or shifts; evaluating or disciplining workers. | A qualified person who reviews each recommendation and can change it, and a notice that AI was used. |
| Money | Personal investment or tax advice; approving or pricing credit; insurance underwriting or claims. | The same review and notice. For licensed advice, OpenAI also expects a licensed professional to be involved. |
| Housing, education, health, benefits, legal | Rental decisions, admissions or grading that decides certification, treatment eligibility or triage, benefits eligibility, legal outcomes. | The same review and notice, plus whatever local law adds. |
| Drafting help for a professional | A lawyer or recruiter using your tool to draft, with their own judgment before anything is sent. | Keep the professional's review visible in the workflow, and do not let the product claim credentials. |
A qualified person must meaningfully review a high-risk AI recommendation, with the authority to change it, and the individual must be told AI was used; covered areas include finance, credit, insurance, housing, employment, education and credentials, healthcare access, public benefits and legal status.
Anthropic Usage PolicyProhibits automation of high-stakes actions in sensitive areas without human review, and tools that give tailored advice requiring a license without appropriate involvement of a licensed professional.
OpenAI Usage PoliciesCopy the self-check
Fill this in for each product or major feature. It takes about 20 minutes and leaves a record you can show a customer, an investor or the provider if they ask.
Worked example: an AI marketing tool
This example is invented to show how the check works. A two-person startup sells an AI marketing assistant built on both providers' APIs. It has four features.
- Drafted LinkedIn and X posts the founder edits and publishes from their own account: no change needed under either policy.
- A "review booster" that writes five-star reviews in customers' voices: falls under Anthropic's fake-review example and OpenAI's deceit rules. The startup removes it and replaces it with a request-a-review email the real customer chooses whether to answer.
- A cheaper in-house model tuned on 40,000 logged GPT answers to replace GPT for short posts: this is training a substitute model. The startup pauses the tuning, keeps the logs for evaluation only and asks OpenAI in writing.
- A public API that lets agencies call "our model", which is GPT behind a thin layer: this gives third parties direct access, so it needs OpenAI's review before November 20. The startup pauses new API signups until it hears back.
What not to assume
- Do not assume approval will come by November 20; OpenAI has not published a timeline.
- Do not treat these policies as the whole contract; your services agreement, platform rules and local law also apply.
- Do not assume a policy that allows a use makes it legal, or that disclosure under one rule satisfies the EU AI Act.
- Do not move a disallowed use to another provider's model without reading that provider's policy.
- Do not rely on summaries, including this one, for edge cases; quote the policy text when you ask the provider.
Where Soloop fits
Soloop does not give legal advice, does not review your agreements with model providers and cannot request approval from OpenAI or Anthropic for you.
According to the Soloop manual, it can research a market with sources, write and publish X posts and replies through a connected account, send outreach through a connected Gmail account and prepare Google Ads search campaigns after setup and payment, with actions waiting for your approval by default. That fits the allowed side of the marketing list above: drafts you review, published under your own account.
Sources, limitations and refresh
OpenAI's Usage Policies and announcement, and Anthropic's Usage Policy and announcement, were read on November 15, 2026. Neither provider publishes a test for edge cases such as bring-your-own-key tools, and OpenAI has not published an approval process or list of approved gateways; this guide marks those points as unknown or inference. The tables, self-check and worked example are editorial guidance, and the example is invented.
This guide will be reviewed by February 15, 2027, or sooner if either provider publishes an approval process or revises these policies.
Usage Policies effective November 20, 2026, and the October 21, 2026 announcement.
OpenAI Usage PoliciesUsage Policy effective November 12, 2026, and the October 8, 2026 announcement.
Anthropic Usage PolicyThe Soloop manual describes research with sources, X posting through a connected account, Gmail outreach and Google Ads search campaigns after setup and payment; approval comes first by default. It offers no legal review.
Soloop product manual