Gemini 3.8 Live Extended Thinking Can Send turnComplete While the Model Is Still Reasoning

Asynchronous function calling takes the tool call off the audio clock. The function runs in the background while the model keeps streaming speech. That is what lets Extended Thinking open with an early verbal cue and then narrate progress instead of going quiet ^[11]. A blocking declaration puts the call back on that clock, so Google's guidance for Extended Thinking sessions is that function declarations carry `"behavior": "NON_BLOCKING"`; the dev.to writeup of the release says blocking tools "fight the architecture" ^[14].

The split shows up in routing. Extended Thinking exposes `thinking_level` at low, medium or high, and Live runs an interleaved profile with no such setting ^[8]. Google's docs say to use Live when turn-taking must feel instant and the task is direct, and Extended Thinking when the agent has to plan, call slow tools, or reason over complex state ^[9]. A product that does both fast lookups and refund adjudication therefore switches endpoints, and the dev.to post recommends reserving Extended Thinking for high-value flows such as refunds, eligibility and multi-step bookings ^[15].

Google lists Live API audio at $0.005 per minute of input and $0.018 per minute of output ^[10]. Output is 3.6 times input per minute ^[18], and progress narration is output. Keep both streams open for a five-minute call and the audio comes to $0.115, so a thousand of those calls is $115 ^[19]. Google footnotes the per-minute figures as converted from token estimates ^[10], which means the rate you actually pay moves with how densely your speech tokenises. The dev.to post argues for instrumenting cost per successful task, on the grounds that cost per session gets the feature killed for the wrong reason ^[15].

Google says Extended Thinking ranks first on Artificial Analysis' Speech-to-Speech leaderboard, a figure the dev.to writeup flags as vendor-reported and directional ^[7]. For that rank to say anything about your agent, the leaderboard's task mix would have to resemble your traffic: similar turn lengths, similar tool latency, and your languages inside its set. The companion claim from Google is mid-conversation switching across more than 97 languages ^[3], and the same post advises treating automatic detection as helpful but not finished ^[17].

Progress speech is now part of the interface, which means someone has to write it, translate it, and cap how long it runs ^[20]. Google also says 3.8 Live takes live camera input in near real time, for receipts, field service and in-app help ^[16]. For those flows the dev.to post suggests confirming the parsed value aloud, as in "I see invoice number... is that correct?", instead of silent OCR ^[21].