Gemini 3.8 text-to-speech models now available on AI Gateway

Vercel has added two new speech synthesis models from Google to AI Gateway: Gemini 3.8 Flash-Lite TTS and Gemini 3.8 Flash TTS. Both accept text input and return generated audio across more than 100 languages, bringing Google's latest text-to-speech generation directly into the AI Gateway routing layer that developers already use to manage model access and observability.

This piece covers what each model does, what changed in the underlying capabilities, and how a developer can start routing TTS requests through AI Gateway today. If you are building multilingual voice interfaces, podcast-style narration pipelines, or any product that needs programmatically controlled speech output at scale, the details below are relevant to your next integration decision.

What shipped in Gemini 3.8 text-to-speech models now

Vercel's AI Gateway now routes requests to two distinct models: Gemini 3.8 Flash-Lite TTS and Gemini 3.8 Flash TTS. The model identifiers follow Vercel's gateway convention and point to the Google-hosted endpoints behind the scenes. Both models share the same core capability set — text in, audio out — but differ in the compute tier they sit on, which affects latency and cost profiles. The Flash-Lite variant targets lighter workloads where speed and economy matter more than maximum expressiveness, while the standard Flash variant is positioned for production use cases that demand richer delivery.

Beyond the two-tier lineup, the models introduce two-speaker dialogue mode, a feature that generates a conversation between two distinct synthetic voices from a single text prompt. This is shipped alongside long-form narration support, meaning the models handle extended content — articles, documentation, course material — without chunking or stitching artifacts that have historically plagued TTS pipelines. The capability lands directly in AI Gateway, so existing gateway customers gain access without switching SDKs or managing separate Google API credentials.

What improved in Gemini 3.8 text-to-speech models now

The 3.8 generation advances on two axes compared with prior Gemini TTS offerings: language coverage and delivery control. The 100-plus-language support represents a significant breadth increase, covering scripts and phoneme sets that earlier versions handled inconsistently. Delivery control — the ability to direct pacing, tone, and emphasis through prompt-level instructions — moves from coarse pitch-and-rate sliders to a more expressive instruction interface. Below is a summary of the key capability shifts based on what Vercel and Google have published.

Because the source does not publish per-character pricing or latency benchmarks, those cells are omitted rather than estimated. Builders should verify current pricing in the Vercel AI Gateway dashboard and run latency benchmarks against their target document lengths before committing to either tier in production.

What you gain from Gemini 3.8 text-to-speech models now

Routing TTS through AI Gateway rather than calling Google's API directly gives teams a single observability plane for both language model and speech model traffic. Costs, token counts, and error rates surface in the same dashboard that already tracks LLM calls, removing the split-instrumentation problem that arises when audio generation sits on a separate billing account. For teams that already use gateway-level caching or rate-limit policies, those controls apply to TTS requests by the same configuration path.

The two-speaker dialogue mode is the capability most likely to change what developers choose to ship. Producing a realistic back-and-forth from a single model call — rather than interleaving two separate single-speaker outputs with manual audio editing — cuts pipeline complexity substantially. Long-form narration support means a 5,000-word article can be converted in a single request, eliminating the segment-boundary glitches that occur when applications split documents and concatenate audio chunks. Multilingual teams gain consistent quality across all supported languages without maintaining per-language model routing logic.

How to get Gemini 3.8 text-to-speech models now

Access runs through Vercel AI Gateway. If you use the Vercel AI SDK, update to the latest release to pick up gateway support for the new model identifiers:

npm install ai@latest

To target the Flash-Lite model in a request routed through AI Gateway, set the model identifier to google/gemini-3.8-flash-lite-tts. For the standard Flash variant, use google/gemini-3.8-flash-tts. In a Next.js or Node project using the AI SDK, pass the model string through your configured gateway provider rather than the bare Google provider to ensure observability and routing policies apply.

```ts
import { createGateway } from 'ai/gateway';

Check the Vercel AI Gateway dashboard to confirm the models appear in your available model list, and verify that your gateway plan covers audio generation endpoints before sending production traffic.

What to watch after Gemini 3.8 text-to-speech models now

The 3.8 label on both models suggests Google is treating this as a versioned, iterative series rather than a one-time release. Developers should watch for a 3.9 or 4.x TTS generation that may close the remaining gaps — particularly around per-voice customization and real-time streaming audio output — that the 3.8 generation does not yet address in the published feature set. Vercel's addition of TTS to AI Gateway also signals a broader platform direction: the gateway is evolving from an LLM router into a multi-modality control plane that covers image, audio, and eventually video generation behind a unified API surface.

On the adoption side, the two-speaker dialogue mode will be worth monitoring in developer community feedback. The implementation details — how voice identity is specified, whether speakers can be named or styled differently, and how the model handles turn boundaries in ambiguous prose — will determine whether it becomes a production staple or remains a demo-friendly novelty. If Google publishes phoneme-level benchmarks or MOS scores for the 100-plus-language set, those numbers will be the clearest signal of where the model holds up against alternatives and where it does not.