GPT Live 1

GPT Live 1 can listen and speak at the same time. BeAI uses it for voice and GPT-5.6 Luna for reasoning and tool selection. Existing tools retain the bot's permissions and organisation scope.

Enable on a DEV bot

Open the bot, choose Edit, then select GPT Live 1 (voice + GPT-5.6 Luna reasoning) in the voice settings. Keep Cedar, check the language, greeting and instructions, save and make a test call on DEV. You need permission to edit the bot. Existing bots retain their model until you change it.

The equivalent API operation is PATCH /api/v1/bots/{id}/voice:

{"ai_model":"gpt-live-1","voice":"cedar"}

The API key's user must have edit permission for that bot and organisation. Selecting this model grants no additional tool access.

Behaviour and limits

Business instructions and tool definitions go to the reasoning backend. The voice model communicates its verified results. Realtime speech speed and turn detection settings do not apply. Interrupting speech does not cancel a business action already executed. Transcript fragments can overlap; display grouping never triggers an action.

The reasoning backend does not receive raw audio. Answering-machine detection based on a beep or other acoustic evidence is not guaranteed; separately test outbound workflows that depend on it. If the OpenAI session cannot start, the call cannot use this model. Check conversation status and errors before retrying.

Costs

Provider voice sessions cost USD 0.05 per minute, billed per second without rounding up to whole minutes. GPT-5.6 Luna tokens cost extra. This is not an all-inclusive bot price.

Conversation costs distinguish voice seconds from backend tokens. Missing final usage or pricing produces a partial estimate. Estimates do not replace OpenAI invoices. BeAI credit charges continue to use the organisation's configured rates.

This release is intended for DEV testing.

Cost details are available to platform administrators in the GUI and through live_cost on GET /api/v1/conversations/{id}.

Prompt, greeting and recording checks

The full bot prompt is supplied to both the voice and reasoning models. The voice also receives the complete generated greeting instructions without truncation. Generated greetings may be rephrased: verify the welcome and interruptions in an actual call.

Use fictional data and disable skills that send messages, transfer calls or change real business data. Test the greeting, a configured fact, a caller correction, and the current time using the read-only get_time tool.

When recording is enabled, wait for call finalization and open the audio in the conversation. The equivalent endpoint is GET /api/v1/conversations/{id}/audio, with the same access permissions. Byte ranges support seeking. Missing files return 404; unauthorized access is denied.

OpenAI voice catalogue

Realtime supports Alloy, Ash, Ballad, Cedar, Coral, Echo, Marin, Sage, Shimmer and Verse. GPT-Live supports these ten voices plus Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder. The additional voices are Live-only. Regional influences do not guarantee a specific French accent.

Select the model and then the voice in the bot settings. The equivalent API is PATCH /api/v1/bots/{id}/voice, for example {"ai_model":"gpt-live-1","voice":"marin"}. Incompatible voices are rejected. Changing only the model selects its default voice when the current voice is incompatible. Existing calls keep their original voice.

Sources: GPT-Live voices, Realtime voices.