Gemini 3.8 Live API: Pricing, Extended Thinking and Migration
Use gemini-3.8-live for low-latency voice dialogue. Choose gemini-3.8-live-extended-thinking when a conversation needs background reasoning and asynchronous tool work. Both are stable API models released September 15, 2026. They are not Gemini 3.8 Flash, and Flash launch prices do not describe their voice-session bill.
Published and source-checked September 16, 2026. This is a documentation-based integration and budgeting guide, not an independently measured latency or quality benchmark.
Quick cost answer
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. 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.
Live versus Extended Thinking
| Decision | Live | Extended Thinking |
|---|---|---|
| Model ID | gemini-3.8-live | gemini-3.8-live-extended-thinking |
| Starting use case | Fast conversational responses | Multi-step reasoning during dialogue |
| Tools | Non-blocking default; blocking compatibility mode | Non-blocking only |
| Thinking configuration | Omit thinking configuration | Low, medium or high; not minimal |
| Input/output limits | 131,072 / 65,536 tokens | 131,072 / 65,536 tokens |
These differences come from the Live model documentation and Extended Thinking migration notes. Do not choose the reasoning variant simply because its name sounds more capable. Measure task completion, interruption recovery and cost using representative conversations.
Pricing by modality
| Billable component | USD per million tokens | Budget implication |
|---|---|---|
| Text input | $0.75 | Instructions and textual context |
| Audio input | $3.00 | Listening input, not just spoken questions |
| Image/video input | $1.00 | Send only frames your task needs |
| Text output, including thinking | $4.50 | Reasoning and transcript costs where applicable |
| Audio output | $12.00 | Generated speech |
The official pricing page also lists free-tier access, different data-use treatment between free and paid tiers, and separately metered Search grounding. These models do not support context caching or Batch API. Do not apply a Flash cache discount, batch discount or its January 2027 rate schedule to Live.
Headline rate versus total session cost
Use actual usage counters for the invoice estimate. As an illustrative token workload, 100,000 audio-input tokens cost $0.30 and 100,000 audio-output tokens cost $1.20. Add 20,000 text-input tokens ($0.015) and 10,000 text-output tokens ($0.045): the example totals $1.56 before tools and transport. This is an arithmetic example, not a claim that a call of a particular duration generates those counts.
For monthly planning, multiply measured cost per completed session by session count, then add provider-specific transport, carrier and tool charges. Keep successful and failed sessions separate. A cheap call that requires a person to repeat the entire task may have a worse completed-task cost than a more expensive successful workflow.
Google's billing guidance says proactive audio is permanently enabled for these models, so listening input is billed while the session listens. Transcript generation can add text-output charges. Session history can also increase later-turn input billing; configure context compression and inspect usage rather than assuming every turn starts with zero history.
Build the voice workflow around durable state
A practical architecture separates the audio connection, application task state and tool executor. The audio layer may be interrupted while a backend operation is still running. Store the task's state outside the transient audio stream so reconnects cannot repeat a booking or lose a pending approval.
Use narrow authenticated tool handlers. A voice instruction is not permission to access another tenant or bypass an approval. Validate tool arguments at the backend, track request IDs and return explicit accepted, completed or failed results. Keep credentials off the client; use Google's documented authentication path for the chosen client/server integration.
Extended Thinking needs interaction-state handling
For the thinking variant, turnComplete: true does not prove the server has finished all reasoning and tool work. Continue receiving messages. Observe interaction_status: IN_PROGRESS means work can continue, while IDLE means the interaction has finished processing.
Design the UI so intermediate speech does not appear as a completed action. Show a pending task separately from the current spoken turn. If the user interrupts, distinguish stopping speech from cancelling application work. On reconnect, reconcile job state before retrying side effects.
Migration checklist
- Select the new model ID rather than treating the old preview as equivalent.
- For standard Live, remove thinking configuration, disabled-proactive-audio settings and removed affective-dialogue options.
- Review tool behavior. Standard Live defaults to non-blocking; Extended Thinking rejects blocking execution.
- For the thinking variant, handle interaction state and use a supported thinking level.
- Use the documented audio response modality; enable audio transcription when your interface needs a transcript.
- Recheck session compression, video submission and billing counters before rollout.
Make this a controlled migration. Keep a rollback configuration, replay a fixed evaluation set and verify behavior before sending all sessions to the replacement. Do not assume a model-string change preserves every prior session option.
Validation is still application work
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. For financial or account-changing tasks, test misheard identifiers, interruptions, ambiguous consent and repeated tool results.
Include quiet input, noisy input, long pauses, tool timeout, disconnect, expired credentials and late results in your test set. Record latency distributions and task outcomes rather than only the fastest demo response. Use privacy-safe evaluation examples and document retention for voice and transcript logs.
Where this fits among other APIs
Choose Gemini 3.8 Flash for its different text/API workload, not as a substitute price row for Live. Compare independent components through the speech API guide when a chained STT, model and TTS architecture is preferable. GPT-Live-1 uses a different session-priced front end with separate backend work. Realtime token-priced APIs belong in a modality-aware comparison, not one universal cost-per-minute ranking.
Our recommendation
Start with standard Live for a fast dialogue task. Move to Extended Thinking only when background reasoning or tool orchestration improves your measured outcome. Keep billing, approvals and durable jobs visible in the design. Neither the release label nor a benchmark ranking establishes that it is the best model for your customers.
FAQ
Is Gemini 3.8 Live the same as Flash?
No. It has a different session API, model identity, output behavior and pricing structure.
Do both Live models have the same published token rates?
Yes, the current Google pricing table groups them together. Actual workload consumption can differ.
Can I use cached or batch discounts?
These Live models do not support context caching or Batch API.
Does a completed audio turn mean a tool job finished?
No. Track interaction and durable application state separately, particularly with Extended Thinking.