Model Terms

1. How models reach you

Sothe is one OpenAI-compatible endpoint in front of models hosted by Amazon Web Services on Amazon Bedrock. When you call /v1/chat/completions, we authenticate your key, check your limits, translate the request into Bedrock's format and send it to the model in a US region. The reply is translated back and billed on the tokens it used.

We are not the model provider, and we do not host, fine-tune or modify the models. The catalogue we list — and whether a model can actually be called — follows what AWS makes available to us, which is re-checked automatically several times an hour.

2. The provider's terms apply to you

Using a model through Sothe does not put us between you and its owner's rules. Each model carries its own licence and acceptable-use policy — Anthropic, OpenAI, Meta, Mistral, Amazon, Google, DeepSeek, Qwen, NVIDIA, Moonshot AI, Z.AI, MiniMax, xAI, Writer and the others in our catalogue all publish their own — and those apply to you directly, alongside AWS's terms for Bedrock and its acceptable-use policy.

Open-weight models come with licences that can impose their own conditions, such as attribution or naming requirements for derivative models. It is your responsibility to read the licence of any model you build on, especially if you redistribute its output or train on it. Where a provider's rules are stricter than our Terms of Service, the stricter rule is the one that binds you.

3. What you send

Sothe does not store your prompts, system prompts, tool definitions, images, audio or video. They are held in memory for the length of the request, forwarded to Bedrock and discarded — see the Privacy Policy. What we keep is the accounting: model, token counts, price, latency and time.

AWS states that Bedrock does not store your inputs or outputs, does not use them to train models, and does not share them with model providers. That is AWS's commitment rather than ours; we route to Bedrock precisely because of it, but we cannot give you a guarantee on another company's behalf.

Where a model supports prompt caching, we may mark the reusable prefix of your request — the system prompt and tool list — so repeated calls cost less. Those fragments live in the provider's cache, scoped to the account making the call, and expire on their own. You can avoid caching by varying that prefix.

Requests have technical ceilings — body size, how many images, audio clips, videos, messages and tools one request may carry — documented under Limits. Media must be sent inline as base64; we will not fetch a URL you supply.

4. What you get back

As between you and Sothe, we claim no ownership of model output and place no additional restrictions on your use of it beyond those in these terms and the provider's licence. We make no promise that output is accurate, original, non-infringing or fit for any purpose.

Models are statistical systems. They invent facts, misquote sources, reproduce biases in their training data and can return text similar or identical to what they return for someone else — so output is not exclusive to you. Anything consequential should be checked by a competent human before it is acted on, and output must never stand alone as medical, legal or financial advice.

5. What you may not generate

In addition to the acceptable use rules in our Terms of Service, you may not use any model on Sothe to produce:

• child sexual abuse material, or any sexual content involving minors;
• intimate imagery of a real person made or shared without their consent;
• material intended to deceive people about a real person's words or actions, or to interfere with an election;
• instructions that meaningfully help someone build weapons, synthesise dangerous agents, or attack computer systems;
• targeted harassment, threats, or content that incites violence against a person or a group;
• automated decisions about a person's credit, employment, housing, insurance or legal status without human review, where the law requires one.

Providers enforce their own rules upstream as well: a request can be refused by the model itself, and repeated attempts can cost us access to that model for every customer. Accounts that use Sothe for the above are closed.

6. Availability, versions and retirement

Model ids are stable strings such as anthropic/claude-opus-5, and carry a provider's version suffix where one exists. A shorter form is accepted when it matches exactly one model we list; if it is ambiguous, the request is refused rather than routed to a model you did not choose.

The catalogue moves. A model may be listed as unavailable because the provider has not enabled it for our account, or as coming soon before a price is published; either can change without notice, and a model can be retired by its provider at any time. GET /v1/models is the authoritative list of what you can call right now. We are not liable for a model becoming unavailable, changing behaviour between versions, or being withdrawn — pin the version you rely on and test before you switch.

7. How you are charged

Prices are per million tokens, listed for each model on the Models page, and follow what our upstream provider charges, so they can change at any time. Input, output and cached tokens are priced separately, and cached reads are cheaper than fresh input.

Before a request runs we reserve the most it could cost and, if a limit of yours cannot cover that, we lower max_tokens to what it can afford rather than refusing outright. You are then charged for the tokens the provider reports. Audio and video are billed by duration, which the file size does not reveal, so the reservation for them is deliberately generous; settlement always uses the provider's own figures.

A request that fails before producing output is not charged. A request interrupted after the model has started generating is charged for what was produced, because the provider charges us for it. See Billing.

8. Using your own provider account

If you connect your own AWS credentials under BYOK, requests run in your AWS account: AWS bills you directly, Sothe charges no credits for those tokens, and your agreement with AWS — not ours — governs the inference. You are responsible for the permissions you grant those credentials and for whatever they are charged.

Your key is always tried first. If AWS rejects it, throttles it, or the model is not enabled in your account, the request either falls back to Sothe's capacity and your credits, or fails — whichever you chose in the dashboard. We store those credentials encrypted and use them for nothing but serving your own requests.

9. Contact

Questions about a specific model, its licence or its availability: [email protected].