Gemini Live and usage

Select a model

Gemini Live 2.5 is the primary default model. Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking remain available as BETA in BeAI. Optimizations preserve explicit model selections and never automatically switch a bot to 3.8 or Extended Thinking.

The same automatic policy applies to Gemini Live bots configured through the GUI or public API. It takes effect on the next call and does not recalculate previous conversations.

Conversation history

In DEV, Gemini Live 2.5, Gemini 3.8 Live standard and Extended Thinking use late context compression: at about 100,000 active context tokens, Google trims the window towards 80,000 tokens. System instructions remain; older turns may leave the model context. This is not a guaranteed summary. Cumulative call usage totals are not the active context threshold.

The policy applies automatically from the next call to bots configured through the GUI or API, with no new permission or setting. It preserves the selected model and business instructions and does not delete conversation history available in BeAI. No local text memory is injected between turns. The bot must reconfirm missing or contradictory information before a write.

Recovering a connection

In DEV, BeAI attempts to resume the Google session after a disconnection or advance expiry notice, using the latest safe checkpoint. The greeting is not restarted. Google Cloud supports replaying unacknowledged messages in their original order. The Developer API used by Gemini 3.8 does not expose these acknowledgements: the latest speech may be missing and the caller may need to repeat it.

Resumption requires temporary context retention at Google and is incompatible with a zero-retention requirement at the provider. BeAI keeps resumption references and replay buffers only in call memory, never in logs.

Recovery is suspended while a tool is pending or its result has not reached a safe checkpoint. A repeated Google action ID is not executed again. A disconnection at that point may end the exchange: check the completed action before requesting it again. Each incident allows at most three recovery attempts, with a bounded replay buffer. Prolonged failures can still interrupt calls.

Speech detection

BeAI uses Google's speech detection and preserves the bot's advanced settings. Where supported by the Google Cloud service, speech start/end events also update call activity. These events do not change your interruption setting. The corresponding setup field is omitted on Google endpoints that reject it.

Without an advanced override, Gemini 3.8 uses HIGH start sensitivity to recognize short replies such as “hello”, with 100 ms of audio retained before detection. End-of-speech detection uses 300 ms of silence, with caller interruption enabled. This is not a guarantee of total response latency.

Existing advanced Gemini settings take priority. Increase start sensitivity if quiet voices are missed. The equivalent API is PATCH /api/v1/bots/{id}/voice, for example {"gemini_live_advanced":{"enabled":true,"start_sensitivity":"high","silence_duration_ms":300}}. Disable the override to restore automatic defaults: {"gemini_live_advanced":{"enabled":false}}. The GUI and API require the same bot edit permissions and organization scope.

When the carrier omits audio packets during silence, BeAI automatically restores that silence for the model. It can then detect the end of your sentence without waiting for your next utterance. This preserves the configured silence duration and does not apply when automatic voice detection is disabled. It does not create speech or transcripts.

A short reply such as “yes” should allow the conversation to continue without waiting for an inactivity warning. The warning checks whether a silent caller is still present; it is not a normal response delay. If the bot only answers after the warning, retain the call reference for investigation instead of lowering the antispam timeout. Actual response time still depends on the model and invoked tools.

Reasoning and tools

Standard Gemini 3.8 Live does not accept a configurable thinking budget, so BeAI sends none. For an explicitly selected Extended Thinking model, BeAI uses LOW, the lowest supported level. This does not eliminate thinking tokens.

BeAI shortens built-in calendar tool descriptions and includes the spoken-announcement rule once. The calendar timezone is supplied by BeAI. Optional business fields such as notes and location remain available, and custom tool restrictions and schemas are preserved.

Calendar results already use short per-call references and a single copy of each event. Diagnostic history retains the full provider result, which is not passed unchanged to the model.

Limits

Resumption reduces interruptions without guaranteeing uninterrupted calls or unlimited retention of every detail. Authentication errors, an unavailable checkpoint or a Google rejection may end the exchange. Keep the conversation reference for diagnosis.

Google counts context processed across turns and tool steps. Usage may grow over the call even with late compression. System instructions and conversation history remain part of the processed context. Savings depend on the scenario; no whole-call percentage is guaranteed. BeAI costs remain estimates, not invoice reconciliation. A Google setup rejection appears as an AI initialization failure, not a successful call.

Instructions sent to the model

For Gemini Live, BeAI places the bot's role, temperament and language first, followed by conversation flow and tools, then guardrails. Your instruction text is preserved. Distinguish information to collect once from topics the caller can revisit. State when each tool should be used, what must be checked and which actions require confirmation. GUI and public API edits use the same permissions.

The bot's default language also provides an input/output transcription hint. It guides recognition without restricting the caller to one language. Select it in bot settings or use default_lang in /api/v1/bots/{id}/prompting. Voice and pronunciation instructions remain applicable. Confirm important names, dates and numbers before an action: transcription errors remain possible.

For Gemini 3.8, BeAI shortens reviewed built-in calendar, time and language templates. Customized technical templates are preserved. Bot and calendar business instructions, prices, capacity and required booking details remain intact.

Pronunciation rules already present in the system prompt are no longer repeated in the greeting request. The greeting text, selected language and recording notice still apply. These automatic changes apply on the next call to bots configured through either the interface or the API.