Supported Models for LLM Gateway

400+ supported model IDs across Claude, GPT, Nova, Llama, Qwen, and more.

​

Overview

The LLM Gateway currently supports 100+ models across Anthropic, OpenAI, and open-weight model families, accessed through Anthropic, OpenAI, or Amazon Bedrock. You can configure which of these models are available to your organization — see Model Access Policies for details.

The table below is live, fetched directly from the gateway’s public model metadata endpoint. You can query that endpoint yourself for the latest supported models and complete metadata.

​

Column reference

Column	What it means
ID	The model ID to pass as model in a request.
Input Cost ($/MTok)	Cost per million input (prompt) tokens.
Output Cost ($/MTok)	Cost per million output (completion) tokens.
Data retention mode	Whether your prompts and responses to this model may be kept: ZDR if they never are, Private if some may be kept for safety review. See Data retention mode below. A dash means the gateway hasn’t reported this for the model yet.
Chat / Responses / Embeddings / Messages	Which of the gateway’s Supported APIs this model can be called through. A checkmark means the model works on that endpoint today.

​

US-only inference

All LLM Gateway inference runs in the United States, for every model and every provider it can fail over to. If you need inference in a different, isolated geography, please contact Aptible Support.

​

Data retention mode

Each model has one of two data retention modes:

• ZDR (zero data retention): none of your prompts or responses sent to this model are kept after the request completes.
• Private: some of your prompts and responses sent to this model may be kept for safety review. That content is never used for training and can’t be read after 30 days. If you need zero data retention, use a model marked ZDR.

Most models are available from more than one provider, and Aptible fails over between them automatically — see Provider Failover below — so a model is ZDR only if every one of those providers is zero data retention. This describes the model providers only; Aptible’s own request and response logs are retained as described in Audit Logging.

A model is Private because of one of these providers:

• Claude Fable 5 and 5.1 on Amazon Bedrock: AWS retains all Fable traffic for up to 30 days for abuse detection, and traffic its classifiers flag may be reviewed by AWS. This data stays within AWS and isn’t shared with the model provider.
• Some GPT models on Amazon Bedrock: AWS retains traffic that its safety classifiers flag for up to 30 days for abuse detection.
• OpenAI models served directly by OpenAI: Aptible’s OpenAI organization runs under OpenAI’s Private Retention with Private Safety Processing. OpenAI keeps traffic its safety systems flag or sample, encrypted, for automated safety review. OpenAI staff can’t read it, and it can’t be decrypted after 30 days. See OpenAI’s Private Safety Processing overview (Appendix A) and data controls guide.

Because of failover, every model published by OpenAI is currently Private, except the open-weight gpt-oss models, which are served only through Amazon Bedrock and are ZDR.

The Data retention mode column above is the source of truth for each model’s mode. If you’re bound by your own zero data retention commitments, use only models marked ZDR, including in a request’s fallbacks.

​

Provider Failover

When a model is available from more than one provider, the LLM Gateway automatically fails over to another available provider if the primary one has an outage. This happens transparently — you don’t need to configure anything or change your request.

​

Client-side fallbacks

You can also specify your own fallback models on a per-request basis by including a fallbacks array in your request body, alongside model:

{
  "model": "claude-opus-5",
  "fallbacks": ["gpt-5", "claude-sonnet-5"],
  "messages": [{"role": "user", "content": "..."}]
}

For example, here’s a complete curl request specifying fallback models:

curl https://llm-gateway.aptible.com/openai/v1/chat/completions \
  -H "Authorization: Bearer $LLM_GATEWAY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "fallbacks": ["gpt-5", "claude-sonnet-5"],
    "messages": [
      {"role": "user", "content": "Tell me a joke about cloud security"}
    ]
  }'

If the primary model is unavailable, the gateway tries each fallback model in order (each with its own automatic provider failover, per the section above) before returning an error.