---
title: >-
  Route, control, and monitor AI model traffic with the new AI Runtime Control
  API
summary: null
url: >-
  https://www.fastly.com/documentation/reference/changes/2026/09/add-ai-runtime-control-api
---

You can now route your applications' requests to AI model providers through Fastly and control them using the new [AI Runtime Control API](https://www.fastly.com/documentation/reference/api/ai-runtime-control/). The gateway works with existing provider SDKs, so you can adopt it by changing only the base URL and credentials, then add per-key rate limits, automatic failover, and usage reporting.

- **Connect providers:** [Create provider connections](https://www.fastly.com/documentation/reference/api/ai-runtime-control/provider-connections/#arc-create-provider-connection) to supply the credentials AI Runtime Control uses to reach a provider on your behalf. Use the provider's own API key or, for Amazon Bedrock, an AWS IAM role. You can also [look up supported providers and models](https://www.fastly.com/documentation/reference/api/ai-runtime-control/providers/#arc-get-supported-providers) and get the [Bedrock setup information](https://www.fastly.com/documentation/reference/api/ai-runtime-control/providers/#arc-get-bedrock-setup) needed to configure a Bedrock connection.
- **Control access:** [Create virtual keys](https://www.fastly.com/documentation/reference/api/ai-runtime-control/virtual-keys/#arc-create-virtual-key) that your applications use to authenticate to the gateway. Each key can have request-per-minute and token-per-minute limits. [Configure a failover policy](https://www.fastly.com/documentation/reference/api/ai-runtime-control/virtual-keys/#arc-update-virtual-key-failover) on a key to automatically retry against alternate provider connections after timeouts, rate limiting, or server errors.
- **Send requests:** Point your existing SDKs at the [AI Runtime Control gateway](https://www.fastly.com/documentation/reference/api/ai-runtime-control/gateway/), which works with the OpenAI Chat Completions and Responses APIs, the Anthropic Messages API, and the Amazon Bedrock Runtime invoke APIs.
- **Monitor usage:** [Get usage metrics](https://www.fastly.com/documentation/reference/api/ai-runtime-control/usage-metrics/#arc-get-usage-metrics) for requests, sessions, and token consumption, filtered by virtual key, provider, or model, with an option to export as CSV. [Get session logs](https://www.fastly.com/documentation/reference/api/ai-runtime-control/sessions/#arc-get-session-logs) with full request and response data for debugging and audits.