Gemini 3.8 Live for the agent loop: does interleaved reasoning actually help?

What it is

Two speech-to-speech models on Google's Live API, released September 15, 2026. Google is launching Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking to make voice interactions more natural, fluid, and intelligent. These models handle complex reasoning, real-time visual context, and background task execution without interrupting your conversation. The plain version is the low-latency default; Extended Thinking adds background reasoning while the conversation keeps going.

The next-work-session test

Scenario: you're building a voice support agent that needs to look up an order, verify a customer, and then decide whether to refund — the classic "please hold while I check" moment that today's voice loops handle by going silent for 4–8 seconds.

The concrete change is asynchronous tool calls. Asynchronous function calling: Async execution (behavior: NON_BLOCKING) is now the default function calling mode. That means your lookup_order call no longer blocks the audio stream — the model can keep talking (or acknowledge the user's follow-up question) while your tool round-trip resolves. If you've been hand-rolling that pattern with filler audio and speculative execution, you can delete a layer of glue code.

The Extended Thinking variant is the interesting one for agent loops that need actual reasoning between tool calls, not just fast reflex responses.

Pricing

Verified via Google's launch coverage and third-party API guides.

• Both models are priced at $0.005/min for audio input and $0.018/min for audio output. Google states this estimate is based on $3/1M input tokens and $12/1M output tokens.
• Google lists audio input at $3 per million tokens and audio output at $12 per million tokens for both models. Text input is $0.75/M; text output, including thinking tokens, is $4.50/M.

Note: those per-minute numbers are Google's estimate, not a flat call price. Those headline rates are not one flat price per minute of an entire call. Listening, generated speech, accumulated context, transcripts and external tools must be considered separately. Budget for the tail — long sessions with growing context are where the bill drifts.

What we'd actually use it for

Narrower than "production voice agent": we'd use it for the reasoning-heavy voice tier of an existing telephony stack, where the current model can't handle multi-step lookups without going quiet. Specifically, the async tool calls plus interleaved reasoning solve a real integration problem — you no longer have to fake presence during tool round-trips.

For text agents, ignore this release. Use Gemini 3.8 Flash or your existing stack. This is only interesting if audio is on the critical path.

Limits

• No self-hosting. They are hosted models, not open weights, so there is no self hosted option. Latency floor is whatever Google's edge gives you.
• Breaking changes from the preview. If you're migrating from gemini-3.1-flash-live-preview, Affective dialogue is removed from the API. Remove any enable_affective_dialog configurations from your code. And Thinking level: thinking_level is not supported for gemini-3.8-live. — you pick the model to pick the thinking behavior.
• Vendor benchmarks are vendor benchmarks. Google claims Gemini 3.8 Live Extended Thinking ranks #1 on Artificial Analysis' Speech to Speech Quality Index with 82.6. It scores 68.6% on τ-Voice, 35.1% on Sierra's τ-Voice-banking, and 97.7% on Big Bench Audio. Treat these as claims until your own eval set says otherwise — especially τ-Voice-banking at 35.1% is a reminder that "leading" is not "solved."
• No structured output enforcement. These Live models do not advertise structured-output enforcement. Validate tool arguments and results yourself. A transcript is not a signed transaction record, and spoken confirmation does not replace backend verification. Your eval harness still has to cover misheard IDs, interruptions, and tool timeouts.
• Video is on by default now. Turn coverage: Defaults to TURN_INCLUDES_AUDIO_ACTIVITY_AND_ALL_VIDEO. Video frames are sent to the model by default, so only send frames when needed to manage context and cost. Easy way to blow up your token bill if you don't gate frames.

Try it if

• You're building or maintaining a voice agent and the current model goes silent during tool calls.
• You need reasoning between turns (banking flows, technical triage) and want to test whether Extended Thinking beats a two-model architecture.
• You already integrate through LiveKit, Pipecat, or Vercel — the plumbing is done.

Skip it if

• Your agent is text-only. This buys you nothing.
• You need open weights, on-prem, or a hard latency SLA you control.
• Your eval discipline isn't in place. The per-minute pricing rewards teams who can measure task completion, not just demo well.

Migration details and the current capability list are in the Gemini 3.8 Live model docs; Google's announcement is on the Google blog.