Overview
Gemini 3.1 Flash Live is Google's highest-quality native audio model to date, released on March 26, 2026, as a preview through the Gemini Live API. Unlike prior Flash models optimized for text throughput and reasoning speed, this release is purpose-built for real-time multimodal voice interactions—processing audio, video, text, and images in a single end-to-end pipeline without intermediate transcription steps.
The model is designed for developers building production voice agents: it handles background noise, supports interruptions mid-sentence, covers 90+ languages, and can execute tool calls during live audio sessions. Verizon, The Home Depot, and LiveKit are among the early adopters deploying it in conversational workflows. It is distinct from Gemini 3.1 Flash-Lite, which targets text-based batch workloads; Flash Live is exclusively for real-time audio and voice applications.
What's New
Native End-to-End Audio Processing
Earlier Gemini Live models already supported native audio capabilities. Gemini 3.1 Flash Live improves latency, reliability, and acoustic understanding over prior native-audio models rather than introducing native audio reasoning for the first time. Gemini 3.1 Flash Live processes audio directly as a native modality—preserving pitch, pace, and tone information throughout inference. This enables the model to recognize acoustic nuance such as hesitation, emphasis, and background sound, producing more contextually appropriate responses without an intermediate transcription layer.
Improved Interruption Handling
Gemini 3.1 Flash Live supports real-time interruption through the Live API's barge-in behavior, allowing user speech to interrupt ongoing generation for more responsive conversations. This capability existed in the Live API earlier, but Google positions 3.1 Flash Live as improving latency and overall dialogue quality.
90+ Language Coverage
Gemini 3.1 Flash Live supports real-time multimodal conversations in over 90 languages. This extends the model's viability for global enterprise voice agents beyond the narrower language sets of previous Live API models. Verify auto-detect language availability for this specific model in the official Live API documentation before relying on it in production.
Improved Function Calling in Live Sessions
Tool use and function calling during real-time audio sessions is significantly improved. The model supports synchronous function calling during live sessions, allowing it to trigger external APIs and retrieve structured data mid-conversation. This makes it viable for AI agent workflows that require live API orchestration—booking lookups, inventory queries, customer data retrieval—while maintaining conversational flow.
thinkingLevel Parameter
Gemini 3.1 introduces the thinkingLevel parameter (replacing the thinkingBudget approach used in prior versions), with four settings: minimal, low, medium, and high. The default is optimized for lowest latency, suitable for most real-time voice interactions. Higher thinking levels improve response quality for complex multi-step queries at the cost of increased latency—developers can tune this per-request to balance response speed against reasoning depth.
Performance Benchmarks
| Benchmark | Gemini 3.1 Flash Live | Notes |
|---|---|---|
| ComplexFuncBench Audio | 90.8% | Multi-step audio function calling with constraints — leads this benchmark |
| Scale AI Audio MultiChallenge | 36.1% | With thinking enabled — leads this benchmark |
ComplexFuncBench Audio evaluates multi-step function calling with structured constraints in a live audio context—the primary use case for voice agents executing tool calls. The 90.8% score represents the state of the art on this task as of the release date.
The model also demonstrates lower latency compared to Gemini 2.5 Flash Native Audio, though exact latency figures depend on thinkingLevel configuration and request complexity.
Compatibility Notes
Supported features in Gemini 3.1 Flash Live:
- Audio generation (native output)
- Function calling / tool use
- Live API (real-time streaming)
- Search grounding
- Thinking (
thinkingLevelparameter) - Text input and output
- Image and video input
Not supported in this release:
- Batch API
- Context caching
- Code execution
- File search
- Image generation
- Structured outputs
- URL context
Token limits:
| Limit | |
|---|---|
| Input token limit | 131,072 tokens |
| Output token limit | 65,536 tokens |
| Knowledge cutoff | January 2025 |
Developers migrating from thinkingBudget in prior versions must update to the thinkingLevel parameter—the older field is not supported in this model.
Availability & Access
Gemini 3.1 Flash Live is available as a preview through:
- Google AI Studio — accessible directly from aistudio.google.com with a Google account; free tier available within AI Studio quota limits
- Gemini API — use model ID
gemini-3.1-flash-live-previewin API requests - Vertex AI — available for enterprise deployments with Vertex AI quota and billing
The model powers the Gemini Live feature in the Gemini consumer app (described by Google as its "biggest upgrade yet" to Gemini Live) and is rolling out globally in Google Search Live. The model is currently available in preview; monitor the official Google AI changelog for updates on stable model IDs.
Pricing & Plans
Gemini 3.1 Flash Live is accessible via the Gemini API and Vertex AI under Google's standard AI voice generator and Live API pricing:
- Free tier: Available within Google AI Studio quota limits for development and testing
- Paid API access: Published Gemini API rates include $0.75/1M text input tokens, $3.00/1M audio input tokens ($0.005/min), $1.00/1M image/video input tokens ($0.002/min), $4.50/1M text output tokens, and $12.00/1M audio output tokens ($0.018/min); verify current rates on the Google AI pricing page as preview pricing may change at GA
- Vertex AI enterprise: Custom pricing and quota available for high-volume production deployments; contact Google Cloud sales for dedicated capacity
As a preview release, pricing and capabilities may change before general availability. Developers building production integrations should verify the current Gemini API pricing page and Live API capability docs before committing to volume.
Best For
- Developers building real-time voice agents for customer experience, support automation, or enterprise workflows that require natural interruption handling
- Teams deploying multilingual AI agent applications across 90+ languages without per-language routing complexity
- Engineers building voice interfaces that need to execute live tool calls—inventory lookups, booking APIs, customer data retrieval—mid-conversation
- Organizations evaluating voice AI infrastructure platforms such as Vapi who want to compare Google's native audio model as the underlying model choice
- Researchers and developers testing real-time audio AI capabilities who need free access through Google AI Studio before committing to production API costs
- Product teams building on Gemini Live or integrating Google Search Live who need the underlying model documentation for performance planning
FAQ
How does Gemini 3.1 Flash Live differ from Gemini 3.1 Flash-Lite?
Gemini 3.1 Flash-Lite is optimized for high-volume text workloads—fast batch processing, reasoning tasks, and UI generation. Gemini 3.1 Flash Live is purpose-built for real-time audio and voice interactions through the Live API. They serve fundamentally different use cases and are not directly substitutable.
What is barge-in and why does it matter for voice AI?
Barge-in means the model stops generating audio output the moment it detects new user speech, and immediately begins processing the new input. Without barge-in, users must wait for the AI to finish speaking before responding—creating the stilted, turn-by-turn dynamic of early voice assistants. Barge-in enables the natural overlapping conversation flow that users expect, significantly improving perceived responsiveness in real-time voice applications.
What is the thinkingLevel parameter and how should I configure it?
thinkingLevel controls how much internal reasoning the model performs before generating a response, with four values: minimal, low, medium, and high. The default is optimized for lowest latency. For simple voice queries or real-time conversational responses, minimal or low is appropriate. For complex multi-step queries or tool orchestration requiring more accuracy, medium or high improves quality at the cost of increased response time. Set it per-request based on query complexity.
Does Gemini 3.1 Flash Live support structured outputs or code execution?
No. Structured outputs, code execution, batch API, context caching, file search, image generation, and URL context are not supported in this release. The model is focused on real-time audio interactions and function calling within live sessions. For use cases requiring structured outputs or code execution, Gemini 3.1 Pro or other Gemini variants are more appropriate.
When will Gemini 3.1 Flash Live reach general availability?
Google has not announced a general availability date as of the March 26, 2026 preview release. Preview model IDs may change at GA. Monitor the Google AI changelog for updates on stable model IDs and pricing finalization.
