Model Usage Policy

Effective Date: July 25, 2026
Last Updated: July 25, 2026

This policy explains which models power our Services, how we handle the third-party providers behind them, and what happens when models change, become unavailable, or are retired.

It forms part of our Terms of Service.

1. How the Platform Works

Our Services are a gateway. When you submit a generation, we route it to a model appropriate for the request, run our own safety and accounting layers around it, store the result, and return it to you.

Some models are operated by us. Most are operated by third-party providers under commercial agreements. In either case, your relationship is with us, and this policy and our Terms govern your use.

2. Model Providers

We work with the following model providers:

| Provider | Used for | | --- | --- | | Black Forest Labs | Image generation and editing | | Google DeepMind | Image generation and editing | | ByteDance | Image generation and editing | | Lightricks | Video generation | | Kling AI | Video generation | | xAI | Video generation | | Sync Labs | Lip-sync | | ElevenLabs | Text-to-speech and voice cloning | | Inferon | Avatar and world generation |

The current catalogue of available models, their capabilities, and their credit cost is shown in the product, which is always the authoritative list. Providers are also listed, with their role and location, in Subprocessors.

3. Model Availability

Model availability is not guaranteed.

Availability may vary by plan, by region, by workload, and over time. A model may be temporarily unavailable due to capacity limits, provider maintenance, or an upstream incident.

We may restrict a model to certain plans, or limit concurrency for a model, in order to keep the platform stable.

4. Provider Substitution

We reserve the right to change the provider or underlying implementation behind a capability.

We may do this to improve quality, reduce cost or latency, respond to a provider outage or policy change, or replace a model that has been retired upstream. Where a substitution materially changes output characteristics for a named model, we will make reasonable efforts to give notice.

A change of provider does not by itself change your rights under our Terms.

5. Model Versions and Deprecation

Providers update their models, and version identifiers change. A model offered under a given name today may be backed by a newer version tomorrow.

Where a provider releases a new version, we may:

Offer it alongside the existing version
Upgrade the existing model in place, where the change is an improvement without significant behavioural difference
Retire the older version once the provider withdraws it

When we retire a model, we aim to give 30 days' notice to affected accounts and to suggest a replacement. Where a provider withdraws a model with less notice than that, or removes it immediately, we may have to act faster, and we will tell you as soon as we can.

Retiring a model does not affect Outputs you have already generated.

6. Credits and Accounting

Generation consumes credits.

Credit cost varies by model, output type, resolution, and duration.
The cost of a generation is shown before you confirm it.
Credit costs may change as provider pricing changes, model efficiency improves, or new models are introduced. We will give notice of increases affecting models you actively use.
Credits are deducted when a generation is accepted for processing.
Credits are restored where a generation fails because of an error on our side or an upstream provider failure.
Credits are not restored because you are dissatisfied with an Output.

Full credit and billing terms are in our Refund & Cancellation Policy.

7. Data Handling by Providers

When your generation is routed to a third-party provider, your prompt and any input files are transmitted to that provider for the time needed to produce the Output.

We select providers that offer non-training or zero-retention handling for API traffic where such terms are available, and we contract for it where we can. Providers process the data under their own terms in addition to our agreement with them.

We do not use your content to train our own general-purpose models. See our Privacy Policy for the full position, and our Biometric Data Policy for how facial and voice data is handled.

8. Safety Filters

Both we and our providers operate safety filters. A generation may be blocked by our systems, by the provider's systems, or by both.

Filters are imperfect in both directions, and a blocked generation is not an accusation. Our approach and the appeal route are described in our Moderation Policy.

Where a provider blocks a request, we may not receive a detailed reason, and we may not be able to explain the decision beyond reporting it.

9. Provider Outages

Upstream providers occasionally fail. When they do, generations using the affected model may queue, slow down, or fail.

In an outage we will attempt to route to an alternative model where a suitable one is available, restore credits consumed by failed generations, and report the disruption as described in Status & Incident Communication.

We are not liable for losses arising from a third-party provider outage. Uptime commitments, where they exist, are set out in Service Levels.

10. Rights in Outputs

Your rights in Outputs are set out in our Terms of Service.

Where an Output is produced by a third-party model, the provider's terms may impose additional conditions on commercial use. We contract for commercial use rights for the models we offer, but we cannot warrant that every Output is free of third-party claims. If you plan a high-stakes commercial use, review the Output independently.

11. Custom and Fine-Tuned Models

Where you request a custom or fine-tuned model, we train it on the material you supply.

The resulting model is provisioned for your use and is not used to serve other customers.
You are responsible for holding the rights to the training material, including any consent required where it depicts a person.
Custom models are retained while your account is active and are deleted on account closure or on request.
Training consumes credits, which are not refundable once training has begun.

12. Changes

We may update this policy as our model catalogue and provider relationships evolve.

The "Last Updated" date indicates the latest revision. Material changes affecting models you use will be communicated to account holders.