Multiple keys and automatic fallback

Connect both an Anthropic and an OpenAI key, and Ochre treats them as primary and backup. If the primary fails on a request, the dispatcher quietly retries against the backup. Your tickets keep moving.

Why bother

LLM providers go down. Rate limits hit during product launches. Maintenance windows happen. If you depend on AI for first response on the chat widget or to triage your queue, a 30-minute provider hiccup is a real problem.

Two keys, one cheap fix.

How fallback decides

A request rolls over to the backup when the primary returns:

• A rate limit error (HTTP 429).
• A 5xx server error.
• A timeout past the provider's normal response window.

It does not fall back on:

• Bad input. If your prompt is broken, both providers will fail the same way.
• Authentication errors. The system surfaces these so you can rotate the key.
• Content policy refusals.

Setup

1. Connect both keys. See BYOK: bring your own Anthropic or OpenAI key.
2. Open AI → Keys.
3. Each connected provider has a default model picker. Whichever provider you draft on first is the primary; the other is the backup.
4. Save.

A common pairing is Claude Sonnet 4.6 as primary and GPT-4o-mini as backup, or the other way around. See Choosing a model.

What gets logged

Every fallback event shows up in two places:

• The conversation receipt notes that fallback was used, which provider was tried first, and why it failed. See AI receipts.
• The receipts feed and dispatch-failures widget on the AI overview surface aggregate fallback rate over time so you can spot patterns.

If your fallback rate spikes above 1 to 2%, something is wrong with your primary key, model access, or quota. Drill into individual receipts for the error reason.

What fallback does not protect against

• Both providers down. Rare, but it happens for short windows. The inbox keeps working; drafting just pauses.
• Bad keys. A revoked or expired primary will fall over to backup correctly, but if your backup is also broken, no AI replies will go out.
• Spend caps hit. When you hit your monthly budget the AI pauses. Fallback does not bypass Spend caps and alerts.

Per-channel keys

You cannot pick a different primary per channel. The primary and backup are workspace-wide. You can pick a different default model per channel, but the keys behind those models are the workspace primary.

Cost implications

Fallback rarely changes your bill in a noticeable way. If your fallback rate is below 2%, the cost difference is rounding error. If it is consistently above that, your primary is unreliable and the bigger problem is reliability, not pricing.

Recommended setup

• Connect both keys, even if you start by using only one.
• Set the cheaper, faster model as backup. Haiku 4.5 or GPT-4o-mini as backup keep fallback events cheap.
• Watch the fallback rate weekly for the first month.

Related

• BYOK: bring your own Anthropic or OpenAI key
• Choosing a model
• AI receipts
• Spend caps and alerts