Two new Gemini live models from Google, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, went live on Vercel's AI Gateway on September 15, 2026. The rollout was announced in a changelog post credited to Vercel engineers Kevin Dawkins, Zachary Chen, and Jerilyn Zheng, and it gives developers access to Google's newest real time audio models through the same routing and billing layer Vercel already uses for dozens of other model providers.

Both models are built for spoken interaction rather than typed chat. According to the changelog, they're aimed at voice assistants, conversational apps, and anything that needs to respond out loud instead of on a screen. That puts them in direct competition with Google's own Gemini Live API, OpenAI's Realtime API, and the growing list of voice infrastructure startups building on top of both.

What's confirmed comes entirely from Vercel's own post. What's missing, so far, is any matching announcement from Google itself, and that gap is worth sitting with before anyone builds a production voice agent on top of it.

Quick Take

  • Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two real time voice models from Google, are now reachable through Vercel's AI Gateway as of September 15, 2026.
  • Developers connect through the AI SDK's realtime WebSocket API, not a REST endpoint, using a short lived token instead of a static API key.
  • Google has not published its own blog post, model card, or Live API documentation naming "Gemini 3.8 Live" as of this writing, so the specs below currently rest on Vercel's changelog alone.

The Details

Per Vercel's changelog, google/gemini-3.8-live supports real time audio, visual grounding, automatic language switching across 97 languages, and background tool calls that run while the conversation keeps going. google/gemini-3.8-live-extended-thinking builds on that with multi step reasoning that runs in parallel with speech, letting the model acknowledge a request and narrate its progress instead of going silent while it works something out.

Access runs through the AI SDK's realtime API. Vercel's quickstart has developers install @ai-sdk/gateway and a WebSocket client, mint a short lived token through gateway.experimental_realtime.getToken(), open a WebSocket connection, and use a model adapter to serialize outgoing events and parse incoming ones. Vercel frames AI Gateway generally as a single API for calling models across providers, tracking usage and cost, and handling retries and failover, plus what the company calls "higher than provider uptime" performance optimizations, a claim that's Vercel's own and not something I've seen independently measured.

One thing the changelog does not include: pricing. That's a departure from Gemini 3.8 Flash's own AI Gateway debut in early September, which shipped with a stated 50% discount through the end of 2026. No comparable number exists yet for either Live model.

Why It Matters

Background tool calling during an open conversation solves a real problem in voice AI. Earlier real time models tend to go quiet while they look something up, which reads as broken to anyone talking to them. If Gemini 3.8 Live genuinely keeps the conversation open while it calls a tool, that's a meaningful gap closed for anyone building support agents or in app assistants.

The bigger pattern is speed. This is the fourth distinct Gemini rollout on AI Gateway since February, following Gemini 3.1 Pro on February 19, Gemini 3.7 Flash on August 13, and Gemini 3.8 Flash in early September. Vercel isn't alone in racing to add more models and faster access, as I covered when Together AI expanded its own fine tuning platform with live metrics and price cuts earlier this month. Third party gateways are increasingly where developers meet a lab's newest model before the lab's own documentation has finished catching up, which is exactly the situation here.

That's also the reason for caution. Google has spent recent months pushing enterprises to look past model of the week hype and toward governance and interoperability. A model shipping on a partner's infrastructure days ahead of the source labeling it themselves is a small irony worth noting, not a reason to assume the model is fake.

What to Watch

Whether Google publishes its own Gemini 3.8 Live model card or Live API listing, which would confirm the 97 language and background tool calling claims independently. Whether pricing shows up on Vercel's model page or Google's own rate card. And whether early developer reports, the kind that show up on Hacker News within days of a launch like this, back up the no interruption claim for Extended Thinking once people actually build with it under real network conditions.

Key Takeaways

  • Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two real time voice models from Google, launched on Vercel's AI Gateway on September 15, 2026.
  • Gemini 3.8 Live handles real time audio, visual grounding, automatic switching across 97 languages, and background tool calls during a live conversation.
  • Gemini 3.8 Live Extended Thinking adds multi step reasoning that runs alongside speech so the model can narrate progress instead of pausing.
  • Google has not yet published independent documentation confirming the Gemini 3.8 Live name, specs, or pricing beyond Vercel's changelog.

FAQ

What is Gemini 3.8 Live?

Gemini 3.8 Live is a real time voice model from Google that Vercel added to its AI Gateway on September 15, 2026. Per Vercel's changelog, it supports real time audio, visual grounding, automatic switching across 97 languages, and background tool calls while a conversation continues.

Is Gemini 3.8 Live the same as Gemini 3.8 Flash?

No. Gemini 3.8 Flash is a separate, earlier model that landed on AI Gateway in early September with a stated 1M token context window and a 50% launch discount. Gemini 3.8 Live is built specifically for real time spoken interaction over a WebSocket connection, not text based chat or agentic coding work.

How much does Gemini 3.8 Live cost?

Vercel's changelog does not list pricing for Gemini 3.8 Live or Gemini 3.8 Live Extended Thinking. AI Gateway's standard model is to pass through provider pricing without a markup, but no specific per token or per minute rate has been published for either model as of this writing.