Models Aren't Infrastructure: What Google's Gemini 3.1 Flash Live Means for Voice Agent Builders

Google just shipped a purpose-built audio-to-audio model for real-time voice agents. It's called Gemini 3.1 Flash Live, and it's good. But a model, no matter how good, is not infrastructure. Here's what that distinction means if you're building voice agents for production.

On March 26, 2026, Google released Gemini 3.1 Flash Live through the Live API in Google AI Studio and Vertex AI. This isn't a general-purpose LLM with a voice wrapper. It's a dedicated audio-to-audio model designed for real-time speech-to-speech interaction.

The capabilities are real. Flash Live processes audio and vision simultaneously, supports tool calling during live voice conversations, filters background noise in real-world environments, and handles 90+ languages. Google also announced production-level partner integrations with LiveKit, Pipecat, and Voximplant for WebRTC and edge routing.

It's a production offering with billing, partner integrations, and enterprise cloud backing.

Google committing a dedicated model to real-time conversational agents confirms something builders in this space already knew: voice AI is a major category, not a feature bolted onto chatbots. That's good for everyone building here, Inworld included. When the largest cloud provider in the world builds a purpose-built model for your market, it removes the "is this real?" question from every sales conversation and investor meeting in the space.

But look at the partner list. Google ships a model, then immediately points developers to LiveKit, Pipecat, and Voximplant for orchestration, WebRTC handling, session management, and edge routing. That's an acknowledgment from Google itself: the model alone isn't enough. You still need the infrastructure around it.

This is the pattern every time a foundation model provider enters a vertical. They build the best model they can, then rely on a network of partners to handle everything the model can't do on its own. OpenAI did the same thing with the Realtime API: great model, but you still need to build the orchestration, session management, and deployment logic yourself. The question for developers isn't "which model is best?" It's "where does the model end and the infrastructure begin?"

A model takes input and produces output. Infrastructure is everything else: how you route between models, how you handle failover when one goes down, how you A/B test different configurations, how you observe what's happening in production, how you deploy across cloud, on-premise, and on-device environments.

Flash Live is a model. A very capable one. But here's where the model-only approach runs into walls.

Flash Live runs on Google Cloud. Only Google Cloud. If the model regresses in a future update, if Google changes pricing, or if you need multi-cloud deployment, you're rewriting your application. With Inworld's Router, you swap between OpenAI, Anthropic, Google, and Mistral models through a single Chat Completions API endpoint, without changing your code. Your application is model-agnostic from day one.

Google's Live API bills per token, per turn, with compounding context window charges. That means you pay for all accumulated tokens every turn. In a 10-minute voice conversation, every new turn includes the full context of every previous turn in the billing calculation. The longer the conversation, the more each subsequent turn costs. Inworld's TTS runs at $5–10 per million characters, and the Runtime is free. At consumer scale, where you might be running thousands of concurrent conversations lasting 10+ minutes each, the pricing gap is 10–50x. (Note: Gemini 3.1 Flash Live is currently free in preview — this comparison reflects the production billing model that will apply once the preview period ends.)

Google's offering is cloud-only. Inworld supports cloud, on-premise, and on-device deployment (demonstrated live at GDC 2025). For regulated industries, gaming, automotive, and edge use cases, on-device isn't a nice-to-have. It's a requirement. Inworld also offers zero data retention mode, SOC2 Type II certification, HIPAA compliance with BAAs, and GDPR compliance.

A note on languages: Google wins on language count, 90+ to Inworld's 15. That's a real gap. Inworld is expanding language support through 2026, but if you need 90+ languages today, that matters for your decision. We'd rather you know that upfront.

The value of any single model depreciates every time a competitor ships something better. Google ships Flash Live today. Tomorrow, OpenAI or Anthropic or a startup nobody's heard of yet ships something faster or cheaper. If your application is welded to one model, you're starting over each time the market moves.

Infrastructure doesn't depreciate the same way. Inworld's Runtime is a C++ graph execution engine where you compose AI agents from nodes (processing tasks) and edges (data flow). You build a graph once, then swap the models running inside it. The Runtime lifecycle (Build, Experiment, Observe, Explore) means you can A/B test graph variants through the Graph Registry without redeploying, monitor production performance through the Portal, and tune models and prompts in Playgrounds.

That's the difference between a demo and a product. A demo works with one model on one cloud. A product works with any model, on any cloud, with the observability and experimentation tooling that production requires.

The Runtime is free. The Router gives you a unified Chat Completions API across hundreds of LLMs. The TTS is ranked #1 on Artificial Analysis with an Elo score of 1,238 on the Max model. And the Realtime API (currently in research preview) follows the OpenAI Realtime protocol with extended customization over WebSocket and WebRTC.

You don't have to pick between Google's model and independent infrastructure. You can use both. Run Gemini through Inworld's Router alongside OpenAI and Anthropic, and let the infrastructure handle routing, failover, and cost optimization automatically.

If you're evaluating Flash Live, evaluate it. It's a strong model and Google's investment in real-time voice confirms the category is here to stay. But evaluate it as a model, not as a complete solution.

Ask yourself three questions before committing your architecture:

What happens when a better model ships in six months? If your answer requires rewriting your application, your architecture has a single point of failure. Building on an infrastructure layer like Inworld's Runtime and Router means you can adopt the next model (from Google or anyone else) without touching your application logic.

What does your cost model look like at 10,000 concurrent conversations? Per-token, per-turn billing with compounding context charges works fine for prototypes. It can break your unit economics at production scale. Run the numbers for your specific use case before locking in.

Where does your application need to run? Cloud-only works for some applications. But gaming, automotive, healthcare, and any use case with strict data residency requirements will eventually need on-premise or on-device options. Building that flexibility in from the start is cheaper than retrofitting it later.

The best voice agents in production a year from now will not be running on a single model from a single provider. They'll be running on infrastructure that lets them pick the best model for every task, swap models without downtime, and deploy wherever their users are.

Yes. Inworld's Router supports Google models (including Gemini) alongside OpenAI, Anthropic, Mistral, and others through a single Chat Completions API endpoint at api.inworld.ai/v1/chat/completions. You specify models in provider/model format (e.g., google/gemini-2.5-flash).

Inworld TTS charges $5–10 per million characters with the Runtime included free. Google's Live API bills per token per turn with compounding context window charges, meaning total cost grows with conversation length. At consumer scale with extended conversations, Inworld's pricing is 10–50x less expensive.

Yes. Inworld demonstrated on-device deployment at GDC 2025 and supports cloud, on-premise, and on-device configurations. Google's Flash Live is cloud-only through Google Cloud.

Inworld's TTS delivers P90 latency under 120ms on the Mini (ultra-fast) model and P90 under 250ms on the Max (flagship) model. Google has not published specific latency benchmarks for Flash Live.

Yes. The Runtime, including the C++ graph execution engine, SDKs (Node.js, Unreal Engine, Unity), CLI tooling, and the experimentation framework, is free. You pay for the AI services you consume through it (LLM calls via Router, TTS calls), not for the infrastructure itself.