Gemini 3.8 Live is GA: voice agents that reason, cheaper than rivals
Two stable Live API models bring mid-conversation reasoning to voice, undercutting GPT-Live-1 and Grok Voice on price while topping their audio benchmarks.
Copy markdown
Two stable models, live in the Gemini API
gemini-3.8-live drives low-latency voice agents; gemini-3.8-live-extended-thinking adds background reasoning without stalling the conversation. Both are listed Stable in Google's model docs and reachable now through the Gemini API, AI Studio and Search Live — no waitlist.
Undercuts GPT-Live-1 and Grok Voice
Reported audio pricing is ~$0.84/hr for the base model and ~$3.50/hr for Extended Thinking, versus $5.83/hr for GPT-Live-1 Astra and $4.80/hr for Grok Voice Think Fast 2.0. On the τ-Voice benchmark it reportedly scores 68.6% against 67.9% and 56.5%.
It reasons mid-call, not between calls
Extended Thinking is tuned to work through higher-complexity, multi-step requests without breaking conversational flow — the gap between a bot that stalls and one that can run a task while you keep talking. Reported speech-to-speech quality of 82.6% edges both rivals.
Build this weekend: a reasoning voice agent for cents
At sub-$1/hr for the base tier, a persistent voice assistant that looks things up and calls your tools shifts from demo to affordable. Point the Live API at gemini-3.8-live, then switch to the Extended Thinking model only for the hard turns to keep the bill down.