> ## Documentation Index
> Fetch the complete documentation index at: https://docs.flexinference.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Backup Providers

You can give us an ordered list of routes for your model with the `provider` array. The first route is the primary and the only one that can run a Flex Race. The rest are backups for the same model. We try them in order when a route fails.

```json theme={null}
{ "start_within": "default", "provider": ["anthropic", "bedrock"] }
```

Leave the array out and we use the model's own direct route.

* `gpt-*` and `o*` run on `openai`, with `foundry` on Azure AI Foundry as the backup.
* `gemini-*` runs on `google`, with `vertex` on Google Vertex AI as the backup.
* `claude-*` runs on `anthropic`, with `bedrock` on Amazon Bedrock as the backup.
* The open-source families run on `cloudflare` or `deepinfra`. Most run on one of them. Two run on both.

A backup changes the route and nothing else. Your model and your tier stay the same, and a backup route always runs a standard call.

We move to a backup on a `429`, a `5xx`, a connection error, or a timeout, and only before the response commits. A chain that used up its routes on faults returns `502 all_routes_failed`. When the last failure was a `429`, you get that `429` back instead.

Every route you name needs its own key. We refuse a missing key before any upstream call, and we never fall through to the next route, because that would put you on a provider you didn't choose.

## Claude on Bedrock

We address every Claude model on Bedrock through a global inference profile, so it can run in any AWS region. A request is not guaranteed to run in the region your Bedrock key names. Bedrock can carry a model and still refuse it for your AWS account, which returns `403 bedrock_model_access_denied`. Grant the model in the Amazon Bedrock console under Model access.

The five newest Claude models need two more one-time steps on the AWS account: `claude-opus-5`, `claude-sonnet-5`, `claude-fable-5`, `claude-opus-4-8`, and `claude-opus-4-7`. Run both with an administrator's credentials. The first command opts the whole AWS account into provider data sharing, Amazon requires it for these models, and it stays on until you turn it off. The Bedrock inference key you added in the dashboard can't run them.

```bash theme={null}
aws bedrock put-account-data-retention --mode provider_data_share
```

```bash theme={null}
aws bedrock list-foundation-model-agreement-offers --model-id anthropic.claude-opus-5

aws bedrock create-foundation-model-agreement \
  --model-id anthropic.claude-opus-5 \
  --offer-token OFFER_TOKEN_FROM_THE_PREVIOUS_COMMAND
```

Until both are done those models return `403 bedrock_model_access_denied`. Amazon requires this, not us.

## GPT on Foundry

The `foundry` route runs GPT through Azure AI Foundry instant access, with no Azure deployment step. Add your resource name and one of its API keys under API, then Provider keys. Create the resource in West US 3, the one instant-access region during the preview.

Azure decides which part of the GPT catalog instant access carries. A model it lacks returns `404 foundry_model_unavailable`.

## Open source on DeepInfra

The `deepinfra` route runs open-source models. Add your DeepInfra token under API, then Provider keys.

DeepInfra sells a flex tier. We don't race it.

The route never tells us when it has read your prompt but hasn't started writing yet. So a race we lose has already paid to read your prompt, then pays again in full. That costs more than a standard call, so we send you standard.

`gemma-4-26b-a4b-it` and `llama-4-scout-17b-16e-instruct` run on both `cloudflare` and `deepinfra`. They go to Workers AI unless you name `deepinfra` first.