A new voice, and calls that end well

A phone voice that reacts like a person, and a call that ends the way people end them. This release also stops Facebook and Instagram’s automated messages from being mistaken for a customer replying.

GPT-Live 1

• A new voice model — GPT-Live 1 joins the voice list with twelve new voices, chosen per teammate. Existing voice teammates are unchanged until you switch.
• Same brain, new voice — the voice model speaks; a separate reasoning model still runs every tool, looks up every record and takes every action. Nothing your teammate could do before is lost.
• Natural interruptions — the voice yields on its own when the caller starts talking, with no artificial cut-off.

Calls that end well

• The goodbye finishes — when your teammate says goodbye or transfers the call, the line stays open until the sentence is done, then hangs up. No more cut-off farewells.
• Hang-ups always register — a goodbye reliably ends the call and writes the call record, instead of leaving the line open.
• Minutes billed from the provider — a call closes on the provider’s confirmed duration, replacing an estimate that over-counted every call by three to seven seconds.
• Nothing dials a dead line — once a call is winding down, no late action can dial, transfer or announce on it.

Meta lead ads

• Meta’s messages read as events — a lead-form card or a form dump shows as a collapsed grey event line (“Meta sent lead form · 4 answers”) that expands to the full text, not as a chat bubble.
• Welcomes survive — Meta’s own messages no longer count as the customer replying, so a lead’s welcome is no longer cancelled by the form that triggered it.
• Honest reply rates — response statistics, triage and the pulse ignore Meta automation rows.
• Draft ready, in place — when your teammate has drafted a reply to a lead’s DM, a Draft ready pill on the event line opens it.

Teammate reliability

• Failed turns retry themselves — a turn that errors retries once after thirty seconds; after a second failure it is set aside and, if a contact is waiting, you get an escalation: try again, or reply by hand.
• Runaway replies are capped — every model call carries an output ceiling, and a reply cut at the cap fails the turn instead of sending a truncated message.
• Stuck threads recovered — idle threads with a message waiting between two minutes and six hours are picked up automatically.

Under the hood

• Admin gains an Output budget card: the cap, cap hits linked to their threads, near-cap completions and per-model output percentiles.
• One shared audio codec replaces three copies; per-utterance transcript timing makes voice latency measurable from a call’s log.