AI models and providers
Which OpenAI and xAI models power each part of Reception AI, how to choose the Realtime model, what connected providers can do, and how to estimate costs.
4 min readUpdated Oct 17, 2026Applies to Reception AI 1.0.0
On this page
Reception AI uses two AI providers by default: OpenAI for live, speech-to-speech voice and xAI for text chat and most other AI features. You supply both keys; usage is billed to your accounts.
Who does what
| Feature | Provider and model |
|---|---|
| Live phone calls (Twilio and Plivo relay, SIP) | OpenAI Realtime: gpt-realtime-2.1 by default |
| Homepage live voice demo | OpenAI Realtime over WebRTC, same model setting |
| Caller transcripts on live calls | OpenAI whisper-1 |
| Workspace chat and the website widget | xAI chat, the receptionist's model (default grok-4.5) |
Turn-based phone path (Twilio outbound, Twilio inbound without APP_URL) |
xAI grok-4-1-fast-non-reasoning for replies |
| Workflow branch decisions | xAI grok-4-1-fast-non-reasoning |
| Widget voice input | xAI speech-to-text |
| Reception AI voices (widget, previews, turn-based calls) | xAI text-to-speech |
| OpenAI voices (widget, previews, turn-based calls) | OpenAI gpt-4o-mini-tts |
| AI builder, instruction rewrites, rehearsals | xAI (grok-4.5) |
Receptionists can choose a chat model under Personality → Advanced · AI model: Grok 4.5 (default), Grok 4.6, Grok 4 or Grok 3 Mini. That choice applies to chat, the widget and the turn-based path, not to live calls.
Realtime models
| Model | Notes |
|---|---|
gpt-realtime-2.1 |
The flagship. The default when REALTIME_MODEL is blank. |
gpt-realtime-2.1-mini |
The smaller model, roughly a third of the price. Test it on your own calls before switching. |
Change it under Configuration → AI → REALTIME_MODEL. It takes effect on the next call (settings refresh within a minute); calls already in progress keep their model.
Note
In 1.0.0, calls arriving through a SIP trunk always use the flagship model regardless of this setting. After switching models, check your OpenAI usage dashboard to confirm which model your calls are billed under.
Live sessions use semantic voice activity detection to decide when the caller has finished speaking, so callers can pause naturally and can interrupt the receptionist.
Is xAI used for live voice?
No. xAI does not handle speech-to-speech calls in Reception AI. Every live, interruptible call goes to OpenAI's Realtime API. xAI powers typed chat, the turn-based phone path (where speech is recognised by the carrier, answered by an xAI model and then read out by text-to-speech), voice previews and the AI helpers.
Connected AI providers
Under Apps, a workspace can connect its own AI provider accounts and attach them to a receptionist. Be aware of their limits in version 1.0.0:
| Connection | Status in 1.0.0 |
|---|---|
| OpenAI, Together, OpenRouter (chat) | Not reliable. The receptionist's Grok model name is sent to the connected provider unchanged, which those providers will usually reject. |
| Anthropic (chat) | Not reliable, for the same reason, and tool results are not carried between turns. |
| ElevenLabs, Cartesia (speech) | Limited to spoken replies in chat. Receptionists can only choose built-in voice names, which these services do not recognise. |
| Deepgram (transcription) | Not used yet; transcription always goes through xAI. |
XAI_API_KEY is required for chat even when another chat provider is connected. For dependable results in 1.0.0, use the platform's OpenAI and xAI keys.
Is GPT-Live-1 supported?
Not yet. Reception AI's live voice is built on OpenAI's Realtime API (gpt-realtime-2.1 and gpt-realtime-2.1-mini). GPT-Live-1 uses a different protocol from the Realtime API, so supporting it means new integration work rather than a setting change.
Estimating costs
Reception AI has no markup on AI or telephony: your providers bill you directly. Prices change often, so this page links to the source rather than quoting numbers that would go stale.
| Cost | Charged by | Driven by | Pricing |
|---|---|---|---|
| Live call audio | OpenAI | Minutes of caller and receptionist audio on the Realtime model | OpenAI API pricing |
| Phone minutes and number rental | Your carrier | Call minutes (inbound and outbound) and a monthly fee per number | Twilio voice pricing, Plivo pricing |
| Chat and turn-based replies | xAI | Tokens in and out per message | xAI models and pricing |
| Speech synthesis | xAI or OpenAI | Characters or audio generated | As above |
| Hosting and database | Your hosts | Server size and storage | Your provider |
A useful way to estimate a live call:
cost per call minute ≈ Realtime audio per minute (OpenAI)
+ carrier inbound per minute (Twilio or Plivo)
monthly fixed cost ≈ numbers x monthly number fee + hostingRealtime audio is the largest variable cost for most installs; the mini model is the main lever for reducing it. Measure a few real calls on your own account before you set prices for customers.
Tracking costs in the admin panel
Admin → Finance records provider costs for each completed live relay call (OpenAI Realtime audio and carrier minutes) and for speech synthesis, using a table of provider rates. The rates are seeded estimates: replace them with the prices on your own invoices. Chat tokens, transcription, SIP calls, outbound calls and monthly number fees are not recorded in 1.0.0, so treat Finance as a lower bound and reconcile against your provider bills. npm run costs:backfill rebuilds cost records for older calls at the current rates.
Tip
Set monthly spending limits in your OpenAI and xAI dashboards, especially on a public demo or a SaaS with self-serve sign-up.