---
title: "Retries"
subtitle: "Retry a failed call on the same model before falling back."
---

When a call fails with an error that can pass, the gateway retries it on the same provider and key. When the tries run out, the request moves to the next provider, key, or [fallback model](/docs/ai-gateway/routing/fallbacks). Retries are on by default.

## Set retries

Open the **Routing** tab of the [Gateway page](https://platform.respan.ai/platform/gateway/traffic?tab=routing). Under **Retries**, turn the switch on and pick how many times to try each model and the first wait. The setting applies to every request in the organization.

<img src={"/docs/images/gateway/retries-light.png"} className="block dark:hidden" alt="Retries set to try each model up to 3 times, with waits from 0.2 seconds" />
<img src={"/docs/images/gateway/retries-dark.png"} className="hidden dark:block" alt="Retries set to try each model up to 3 times, with waits from 0.2 seconds" />

| Setting | Default | Detail |
|---|---|---|
| Switch | On | Off tries each model once. Fallbacks still run. |
| Try each model up to | 3 times | Tries per model, the first one included. 1 to 3. |
| With waits from | 0.2 seconds | The first wait, 0.2 to 2 seconds. Each wait doubles the one before, up to 60 seconds. |

Retries apply to chat requests. Images, audio, embeddings, and other requests are sent once per attempt.

## What is retried

| Failure | Retried |
|---|---|
| `408`, `425`, `429`, `500`, `502`, `503`, `504` | Yes |
| Connection failed before any response | Yes |
| Other `4xx`, such as `400` or `404` | No, moves to the next attempt |
| Input larger than the context window, `401` from your own key | No, ends the request |
| Refused by your [guardrails](/docs/ai-gateway/routing/guardrails) or [limits](/docs/ai-gateway/routing/limits) | No, ends the request |

Each model in a fallback gets its own tries, so the worst-case time grows with both. Lower the count for latency-sensitive traffic. A provider call times out after 600 seconds, and a stream that has started isn't retried.

## Logs

Every try is a row in the request's trace. The **Retries and fallback** chart on the Gateway page's **Overview** counts requests that were retried or sent to a fallback model.