Terms of Service

1. About these terms

1.1 These terms are an agreement between INVERGENT S.A., Sos. Pipera 43, Bucharest, Romania, company number RO31021453 ("Surogate", "we", "us") and you. They cover your use of llm.buzz: the model API at https://api.llm.buzz/api/v1, the console at https://app.llm.buzz, and the public website at https://llm.buzz (together, the "service").

1.2 You accept these terms when you create an account or use the service. If you use the service for a company or another organization, you confirm that you may bind it, and "you" means that organization.

1.3 Our Privacy Policy explains how we handle personal data. It forms part of these terms.

1.4 These terms cover your use of llm.buzz as a customer. They do not cover selling model access through llm.buzz as a provider: we agree that with each provider directly.

2. Definitions

• Account: your console account. Your first sign-in creates it.
• API key: a key issued to your account. An inference key calls models. A management key calls only the account routes (/api/v1/key, /api/v1/credits and /api/v1/keys).
• Model: a model name in the llm.buzz catalog, such as author/model.
• Provider: a company whose endpoint serves a model. Cloud providers are model companies and cloud platforms that we hold accounts with. Independent providers are companies we approved through the llm.buzz provider marketplace.
• Request: one call you send to the API. An attempt is one try at one provider. A request can have several attempts when we move it to another provider.
• Content: what you send in a request (messages, system prompts, images, files, tool definitions and results, and parameters) and what the model returns ("output").
• Credit: the prepaid US dollar balance of your account.

3. Eligibility and accounts

3.1 You must be at least 18 years old and able to make a binding contract. [OWNER: say whether the service is for business use only, or also for consumers.]

3.2 You sign in through Firebase Authentication, with an email address and password, with Google, or with GitHub. An email and password account must verify its address before its first sign-in. Keep your sign-in method secure. You are responsible for what happens in your account.

3.3 Give accurate information and keep your email address current. We send notices about your account, payments and these terms to it.

3.4 The console keeps you signed in with a session cookie for 30 days. Sign out on a shared computer.

4. API keys

4.1 An API key authenticates your requests. The console shows a new key once. We store only a fingerprint of it, so we cannot show it to you again. If you lose a key, create a new one and revoke the old one.

4.2 Keep your keys secret. Do not publish them in code, and do not ship them unprotected in apps that run on other people's devices. You are responsible for all use of your keys, including use by people you give them to and use after a key leaks, until you revoke it.

4.3 In the console you can name, disable, enable and revoke keys, and give an inference key a spending limit per day, week, month or for its whole life. We refuse new requests on a key whose spend has reached its limit. Requests already running when the limit is reached still finish and are charged.

4.4 The console's Playground runs on your console session, under one internal key for your account that your browser never receives. Playground requests are authorized, charged and recorded like any other request.

5. The service

5.1 llm.buzz offers one API, compatible with the OpenAI, Anthropic and OpenRouter APIs, in front of many models and providers. For each request we choose among the providers of the model you name, using the model's routing preset (for example by price or speed), provider health, and the preferences you set in the request. If a provider fails or is busy, we may send the request to another provider of the same model. We never answer a request with a different model than the one you named. Each response names the model and the provider that served it (the x-gateway-canonical-model and x-gateway-provider headers).

5.2 You can steer routing for each request. provider.order, provider.only and provider.ignore name the providers to prefer, to allow or to skip. provider.allow_fallbacks: false stops us moving the request to another provider. provider.zdr and provider.data_collection limit the request to providers with a data policy you require (section 8.5). The API reference describes these fields.

5.3 Providers of one model can differ. They can run it at different precisions, with different limits and prices. The model pages show each provider's prices, limits, location and declared data policy, and each independent provider's declared precision.

5.4 The catalog changes. We may add, change, pause or remove models and providers at any time. Providers may change or withdraw what they serve. When a model has no provider that can serve it, the API answers with an error.

5.5 We do not promise any level of availability, latency or throughput unless we agree it with you in writing. Performance figures on the website and in the console describe the past.

5.6 We may set limits that protect the service, for example on request size (the API refuses a body larger than 64 MiB), on concurrent requests or on request rates.

6. Providers and models

6.1 Third-party providers process your requests. Each provider has a data policy: whether it trains on prompts, how many days it keeps them, and whether it is EU-resident. We show each provider's declared policy on the model pages, in the console and in the API's endpoints listing. A provider without a declaration is shown as not declared.

6.2 Independent providers make their own declarations. For cloud providers we make the declaration, based on the provider's published terms and the settings of our account with it. We rely on these declarations. We do not audit providers' systems.

6.3 Some providers are routers themselves, for example OpenRouter. Such a provider may pass your request on to a provider it selects, under its own terms and policies.

6.4 Many models come with terms from their developer or provider, such as a license for an open-weights model or a usage policy. When you use a model you must follow the terms that apply to that use. [OWNER: decide whether the model pages link each model's terms.]

6.5 Providers are independent of us. Except where these terms or the law say otherwise, we are not responsible for what a provider does with the content it receives. [COUNSEL: confirm this allocation.]

7. Credit, prices and payment

7.1 Prepaid credit. You pay for the service in advance with credit in US dollars. We charge each attempt against your balance when it finishes. When your balance is at or below zero, we refuse requests to priced models with HTTP 402 until you add credit. A model whose providers are all free can still be used at a balance of zero or below.

7.2 Prices. Each provider of each model has prices per million tokens for input, output, cached input, cache writes and reasoning, where they apply. An audio model is priced by its input only: a speech-to-text model per million seconds of input audio, a text-to-speech model per million input characters. You pay the provider's price adjusted by our rate. The model pages, the console and the models API show the prices in force. A request that asks for a web search also pays for each search that runs.

7.3 Price at acceptance. A request is charged at the prices and the rate in force when we accepted it. A later price change never reprices it.

7.4 What a request costs. We count tokens from the usage the provider reports. If we move a request to another provider, the attempts that did not produce the answer cost you nothing. A failed last attempt is charged for the usage its provider reports, if any. If a provider reports no usage, or a stream ends before we can measure it, we charge the amount we reserved for that attempt, which is its maximum cost at the prices in force.

7.5 Holds and negative balances. While your balance is below a threshold we set (today $100), we hold the maximum cost of each running attempt until it finishes. We refuse an attempt that could cost more than your available credit. Above that threshold we do not hold amounts, so long or parallel requests can take your balance below zero. You owe a negative balance, and we take it from the next credit you add.

7.6 Buying credit. You buy credit on the console's Credits page through Stripe Checkout, by card, from $5 to $25,000 per purchase. We add a service fee to each purchase. Today it is 5.5% of the credit, at least $0.80, rounded up to the cent. The Credits page shows the fee and the total before you pay. Stripe processes the payment. We never receive your full card number. Your credit appears once Stripe confirms the payment, usually within seconds.

7.7 Auto top-up. You can save a card at checkout and turn on auto top-up. When your balance falls below the threshold you choose ($2 to $25,000), we charge your saved card the amount you choose ($5 to $25,000) plus the service fee. We start at most 5 auto top-ups per account per UTC day. Auto top-up turns itself off when a charge is declined or needs your bank's confirmation, and the Credits page says why. You can turn it off at any time. By turning it on you authorize these charges.

7.8 Refunds. You can refund a purchase or an auto top-up from the Credits page for 24 hours after the payment. You can refund only credit you have not spent and that running requests do not hold. The service fee is not refunded. After 24 hours, purchased credit is not refundable, except [OWNER AND COUNSEL: where the law requires, and what happens to unused credit when an account closes].

7.9 Credit we add. We may add credit to your account, for example as a correction or a promotion. Credit we add is never refunded to a card.

7.10 Nature of credit. Credit is a prepayment for the service. It has no cash value except through the refunds in section 7.8, and you cannot transfer it to another account. [OWNER: confirm that credit does not expire. The product has no expiry today.]

7.11 Invoices and taxes. Stripe issues invoices for purchases and receipts for auto top-ups. Manage billing on the Credits page opens Stripe's customer portal, with your invoices, receipts and payment methods. [COUNSEL AND OWNER: tax treatment. Stripe Checkout does not calculate tax in the current setup; say whether prices include or exclude VAT and sales taxes.]

7.12 Disputes and chargebacks. If you think a charge is wrong, write to us within [DISPUTE WINDOW] (section 17). If you open a chargeback, we may remove the related credit and suspend the account until it is resolved.

7.13 Price changes. Cloud provider prices can change at any time. When an independent provider raises a price, the model page shows the coming price and its date before it applies. We may change the service fee or our rate with 30 days' notice. A change never applies to a request we already accepted or to credit you already bought.

8. Your content

8.1 You keep your rights. As between you and us, you own your content, including outputs. We claim no ownership of it.

8.2 Permission to process. You allow us to process your content to run the service: to send it to the providers that serve your requests, to return their outputs, to keep it as section 8.4 and the Privacy Policy describe, and to keep the service secure. This permission ends for each item of content when it is deleted under the Privacy Policy.

8.3 Your responsibility. You need every right and permission for the content you send, including for personal data in it. Do not send content that you may not share with the providers your request can reach. If your content contains personal data, you are responsible for having a legal basis and for informing the people concerned. [COUNSEL: whether a data processing agreement is offered, and Surogate's role for content.]

8.4 Prompt logging. Unless a request asks for zero retention (section 8.5), we keep the content of completed Chat Completions, Responses and Anthropic Messages requests ("logged prompts") for 30 days, then delete it. Prompt logging is on by default. You can turn it off for your account on the console's Prompt logs page, where you can also read your logged prompts and delete them. Turning it off keeps later requests out and deletes nothing already logged. [OWNER: confirm the purposes for which logged prompts are used.] The Privacy Policy gives the details.

8.5 Zero retention per request. A request that sends "provider": {"zdr": true} goes only to providers that declare they keep no prompts. A request that sends "provider": {"data_collection": "deny"} goes only to providers that declare no training and no retention. Either one also keeps the request out of prompt logging. You can also set your account's data policy to "no training and no retention" on the Prompt logs page: every request of your account then goes only to providers that declare both, and none is logged. If no provider of the model qualifies, the request fails and nothing is sent to a provider.

8.6 Short-term stores. To answer retries and continue conversations, the service keeps some requests for up to 24 hours: a request sent with an Idempotency-Key, and each Responses turn unless you send store: false. These are not prompt logs. They are kept even when a request asks for zero retention.

8.7 Outputs. Outputs come from third-party models. They can be wrong, incomplete or offensive, and they can resemble outputs given to others. Check outputs before you rely on them, in particular for legal, medical, financial or safety decisions. You are responsible for how you use outputs.

8.8 Usage records. We keep records of every request without its content: tokens, cost, timing, states, the provider that served it and the calling app. We use them for billing, reporting, security and running the service. We may publish statistics built from them only as aggregates that do not identify you (Privacy Policy, section 8).

9. Acceptable use

9.1 You must not use the service to:

• break the law, or help someone else do so;
• create, request or share child sexual abuse material, or any content that sexualizes minors;
• build or spread malware, attack systems, send spam or commit fraud;
• infringe other people's rights, including privacy and intellectual property rights;
• break the usage policies of the model developers and providers that apply to the model you use;
• try to learn about other customers' content or activity, for example by timing responses from prompt caches that providers share between customers;
• get around keys, spending limits, credit checks, or the routing and data policy controls, or use another person's account or key without permission;
• overload the API, the console or the public website, including by crawling at unreasonable rates;
• test the security of the service without our written permission. Ask us first on our Discord server (section 17.6).

9.2 You may build products on the service and let others use them. You are responsible for their use of the service as if it were your own.

9.3 We may refuse, limit or stop requests that we reasonably believe break this section.

10. Suspension and closing your account

10.1 You may stop using the service at any time. You can delete your account on the console's Settings page. Any credit left in the account is lost when you delete it, and a negative balance stays owed. The Privacy Policy says what deleting removes and what we keep.

10.2 We may suspend or disable your account or keys if you break these terms, if a payment fails or is reversed, if your use puts the service, providers or other people at risk, or if the law requires it. We do this without notice when we need to act quickly. Disabling an account ends its console sessions, revokes its keys and turns its auto top-up off. Its balance stays recorded.

10.3 We may end these terms for any reason with 30 days' notice. [OWNER: whether unused purchased credit is refunded when we end the terms without cause.]

10.4 Sections 7 (for amounts owed), 8, 11, 12, 13, 14, 15 and 16 survive the end of these terms.

11. Changes

11.1 We may change, add or remove features of the service.

11.2 We may change these terms. We will publish the new version with its date. For a material change we will also email you at least 30 days before it applies, except for changes the law requires, which apply when the law says. If you keep using the service after a change applies, you accept it. If you do not agree, stop using the service before then.

12. Our rights and feedback

12.1 The service, its software, the llm.buzz and Surogate names and logos, and the website belong to Surogate or its licensors. These terms give you a limited, non-exclusive, non-transferable right to use the service while they apply.

12.2 If you send us suggestions or feedback, we may use them without any obligation to you.

13. Disclaimers

13.1 The service is provided "as is" and "as available". To the extent the law allows, we give no warranty that it will be uninterrupted, error-free or fit for a particular purpose, or that outputs will be accurate.

13.2 [COUNSEL: consumer law carve-outs if consumers may use the service.]

14. Limitation of liability

14.1 To the extent the law allows, neither side is liable for indirect or consequential losses, lost profits, lost revenue or lost data.

14.2 To the extent the law allows, our total liability under these terms and for the service is zero.

14.3 Nothing in these terms excludes or limits liability that the law does not allow to be excluded or limited, such as liability for fraud, for damage caused intentionally or by gross negligence, and for harm to a person's physical or mental integrity or health. [COUNSEL: confirm under Romanian law.]

15. Indemnity

You will defend and indemnify Surogate against third-party claims arising from your content, from your use of outputs, or from your breach of section 9. [COUNSEL: scope, and whether it applies to consumers.]

16. Governing law and disputes

16.1 These terms are governed by the laws of Romania.

16.2 The courts of Romania have exclusive jurisdiction. [COUNSEL: arbitration, class action waiver or consumer forum rules, if any.]

16.3 Before starting a claim, write to us (section 17) and give us 30 days to try to resolve it.

17. General

17.1 These terms, with the Privacy Policy, are the whole agreement between you and us about the service, unless we sign another agreement with you.

17.2 You may not transfer these terms without our written consent. We may transfer them to a successor to the llm.buzz business.

17.3 If a provision is unenforceable, the rest stays in effect. If we do not enforce a provision, we do not waive it.

17.4 Neither side is liable for delays caused by events outside its reasonable control.

17.5 You agree that we may give you notices by email or in the console.

17.6 Contact:

• sales and general questions: sales@surogate.ai;
• legal notices, privacy requests and security reports: our Discord server, https://discord.gg/eVTq93BrQ4;
• post: INVERGENT S.A., Sos. Pipera 43, Bucharest, Romania.