AI models & your own keys

Every brand picks the language model behind its assistant. Out of the box you run on the platform default — nothing to configure, usage on the platform meter. Add your own provider key and you can pin any model in that provider's catalog, from a handful of Claude models to OpenRouter's 300+.

1. Pick a model for a brand

Go to Customize & Deploy → AI agent. The AI agent card at the top has two controls:

• Provider — every provider you hold a working key for, plus Platform default. Switching provider pre-selects that provider's first available model.
• Language model — the models that provider offers on your key. This is a type-to-search box, not a plain dropdown: with a handful of models the list opens as soon as you click; with a big catalog it tells you how many match and asks you to keep typing until fewer than 8 do. Every word you type has to appear somewhere in the model's name or id, so deep flash finds DeepSeek V4 Flash without knowing the exact id.

Picking a concrete model fires a quick live test against the provider — you'll see Testing model... and then Tested OK, or the provider's error if the model can't answer on your key. Save when you're happy.

2. Platform meter, or your own key

There are two ways to pay for model usage, and the picker reflects which one you're on:

• Platform meter (default). Your assistant runs on Quincer AI's shared provider keys. You don't add anything, and usage is metered on your Quincer plan. On the shared keys you get the Platform default model — individual models aren't pickable, because that choice belongs to whoever pays the provider bill.
• Your own key. Add a provider's API key and that provider's full catalog appears in the picker for your brands. Usage on your key is billed by the provider to your account, and the rate limits are yours, not shared.

Keys live in Settings → Workspace, in the AI API keys card (“Add your own API keys to unlock additional models”). There's a field per provider — OpenAI, Anthropic, Kindo, X-AI (Grok), Google Gemini, Baseten, OpenRouter and Together.ai — and one Save API keys button. When a provider is currently covered by the shared platform key, the field says so — naming that provider: “Currently using Quincer's shared OpenRouter key. Add your own to bill usage to your account and control the rate limits.”

Every key is checked when you save it: we ask the provider to list its models on that key and report back — Verified with a model count, rejected (the key is saved but nothing will run on it — replace it), or couldn't reach the provider (usually temporary). Saving several keys back-to-back can skip the re-check for a few seconds — save again if you want that one checked too. The same verdicts appear on the Customize & Deploy picker, so a rejected key is never a silent gap in the Provider list.

3. OpenRouter — one key, 300+ models

OpenRouter is a router: one key, and its catalog spans hundreds of models across many vendors — the full list appears in your picker (300+ chat models at the time of writing), which is exactly why the model box is typeable. Three things happen automatically:

• Prices come with the catalog. OpenRouter publishes per-token rates alongside each model, and we import them — so usage on a freshly-picked OpenRouter model meters at the provider's real published rate, not at zero.
• Routing aliases are filtered out. Variant ids like :free and :nitro are routing aliases of models already in the list, so they're excluded rather than doubling the picker.
• A no-training constraint rides every request. Some of OpenRouter's upstream providers reserve the right to train on what they receive, and the choice is per-request. We stamp data_collection: "deny" onto every OpenRouter request at the transport layer, so your conversations only ever route to upstream providers that do not collect or train on prompt data — and OpenRouter's automatic failover still works, but only within the set of providers that satisfy that constraint. No call can opt out of this.

4. Together.ai — the open-weight lineup

Together.ai hosts open-weight models — Llama, Qwen, DeepSeek, Mixtral and friends — on its own inference. Add a Together.ai key and its chat catalog joins your picker. Together's listing also includes embeddings, rerankers, image and moderation models; those are filtered out, so everything you can pick can actually hold a conversation.

5. Voice: choose the OpenAI realtime model

Voice has its own model setting, separate from text chat. If your default voice provider is OpenAI Realtime, you can choose which realtime model answers calls — under Settings → Voice, in the Call behavior card:

1. Set Default voice provider to OpenAI Realtime (gpt-realtime). An OpenAI realtime model field appears below it.
2. Type any model id — it's free text, so you can adopt a new realtime release the day it ships, without waiting for us. The field commits when you click away.
3. Leave it blank to stay on the platform default (gpt-realtime). Clearing the field puts you back on the default.

The model you pick is the model that answers — the same value drives the call connection, the live session, and the per-call cost report, so billing always reflects the model that actually spoke. Two cautions, both shown inline in the dashboard:

• It must be a realtime speech-to-speech model. A model id OpenAI doesn't recognise fails when the call connects.
• Check pricing for a non-default model. Voice cost is looked up by model id — if we hold no price for the id you typed, calls still work but their cost reports as $0 until a price is added.

Voice keys, silence timeout, call length, and the per-persona voice picker are covered in Voice chat.

Troubleshooting