Service-Specific Terms

Different model channels are backed by different providers and licenses. This document sets out, item by item, the additional terms for each channel, the differences in data handling, and the special provisions for beta features.

Effective 2026-07-01Last updated 2026-07-05Version v2.2

This document supplements the Terms of Service. When you call a model through Compass, you are bound not only by our platform terms but also by the applicable terms of that model’s provider; in the event of a conflict, the provider’s terms prevail with respect to the use of that model.

1. Upstream Providers and Applicable Terms

Provider / channel

Additional provisions

OpenAI family

Subject to the OpenAI business terms and usage policies; the enterprise channel is zero data retention, and requests are not used for upstream training.

Anthropic Claude family

Subject to the Anthropic commercial terms of service and acceptable use policy; zero retention by default, with safety filtering performed upstream.

Google Gemini family

Subject to the Google Cloud terms of service and the Generative AI additional terms; some regions are restricted by provider policy.

DeepSeek family

Subject to the provider’s open platform agreement; both mainland China and international channels are offered, and the international channel does not route data back to the mainland.

xAI Grok family

Subject to the xAI developer terms; real-time retrieval capabilities carry separate citation and attribution requirements.

Meta Llama weight models

Governed by the Llama Community License; services exceeding 700 million monthly active users must obtain a license from Meta directly, and products must retain the "Built with Llama" attribution.

Mistral / Qwen weight models

Governed by Apache 2.0 and the Tongyi Qianwen license respectively; license and copyright notices must be retained for commercial use and redistribution.

Enterprise private channels

Governed by the executed enterprise contract, which may specify dedicated rate limits, data residency regions, and a higher SLA.

2. Differences in Data Handling

On the gateway side we never retain request bodies, but retention policies vary across upstream channels. The model detail page in the console indicates the retention tier of each channel; when handling sensitive data, prefer zero retention channels.

Retention tier

Meaning

Zero retention

The upstream provider does not persist requests or responses, does not use them for training, and performs no human review.

Short-term safety retention

The upstream provider retains data for up to 30 days for abuse detection, with human review only when risk controls are triggered, and no use for training.

Self-hosted weights

The model runs on our compute or that of a cloud vendor; data does not leave that region, and there is no third-party retention.

Teams that need to execute a Data Processing Agreement (DPA) or designate a data residency region should contact [email protected]; we provide a standard DPA and an EU Standard Contractual Clauses annex.

3. Use of Output from Open-Weight Models

Some open-source licenses require attribution of the model source in your product; comply with the applicable license.
Using the output of certain models to train competing models may violate their licenses; verify before commercial use.
The slugs and quantization details we offer are stated on each model detail page, to help you assess license and quality differences.

4. Free and Trial Channels

Channels marked as free typically draw on provider promotional quotas; their availability and rate limits may change at any time, and no SLA applies.
Trial credits may not be used on critical paths in production; we may adjust the range of models available to them without notice.
Free channels may exhibit higher queuing latency; use paid channels for latency-sensitive workloads.

5. Beta Features

APIs and features marked Beta are provided as is, may change or be withdrawn, and carry no availability commitment.
Billing rules during a Beta period are stated separately in the feature documentation and may be adjusted at general availability.
We may collect anonymized call statistics to improve the feature, but these never include request bodies.

6. Routing and Failover

A single slug may be backed by multiple upstream channels. Under the default policy, a failed request automatically fails over to a backup channel, which may introduce provider-level differences (such as safety filtering strictness, context limits, and retention tier). If you need to pin a single provider, specify the channel in the request header or disable automatic failover in the console, accepting the corresponding reduction in availability.

7. Order of Precedence

Documents apply in the following order: enterprise contract → these Service-Specific Terms → the Terms of Service → technical notes on the documentation site. Technical documentation is for reference only and creates no contractual obligations.

PreviousSupported Regions NextPrivacy Policy