SunsettingAI Audio

Gemini TTS previews (gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts, gemini-3.1-flash-tts-preview)

by Google

Announced

—

Complete

2026-11-17

Refund window closes

—

Refund status

No refund applicable

Google's Gemini API deprecations page lists November 17, 2026 as the shutdown date for all three preview text-to-speech models: `gemini-2.5-flash-preview-tts` and `gemini-2.5-pro-preview-tts` (both released May 20, 2025) and `gemini-3.1-flash-tts-preview` (released February 26, 2026). Each row names the same replacement: `gemini-3.8-flash-tts` or `gemini-3.8-flash-lite-tts`, which became generally available on September 22, 2026. Google says the listed dates are the earliest possible shutdown dates and that it will confirm the exact date with advance notice. The replacements are cheaper per second of audio until December 31, 2026, then their listed prices double on January 1, 2027. They also change the request format: stage directions typed into the text are now read aloud as part of the transcript, and unary calls return WAV instead of raw PCM. For a pipeline that generates voiceover for AI video, both changes can break output without throwing an error.

Refund flow

1
Not applicable. These are pay-per-token API models with no pre-paid balance tied to a model string, so there is no refund category to open.
2
If you are billed for one of these three model IDs after it is shut down, treat it as a billing error and raise it through Google Cloud billing support with the invoice line, rather than as a refund request.
3
Re-budget before you switch. Paid tier prices per 1M tokens listed on Google's Gemini API pricing page on October 11, 2026 (text input / audio output): gemini-2.5-flash-preview-tts $0.50 / $10, gemini-2.5-pro-preview-tts $1 / $20, gemini-3.1-flash-tts-preview $1 / $20. Replacements: gemini-3.8-flash-tts $0.50 / $9 through December 31, 2026 and $1 / $18 from January 1, 2027; gemini-3.8-flash-lite-tts $0.50 / $6 through December 31, 2026 and $1 / $12 from January 1, 2027.
4
In audio time: Google lists the 3.8 models at $0.00225 (Flash) and $0.0015 (Flash-Lite) per 10 seconds of audio through December 31, 2026, doubling to $0.0045 and $0.003 in 2027. Those figures work out to 25 audio tokens per second. At that same rate, gemini-2.5-flash-preview-tts costs $0.0025 per 10 seconds, and gemini-3.1-flash-tts-preview and gemini-2.5-pro-preview-tts cost $0.005.
5
What that means per migration path, on the 2027 prices: gemini-2.5-flash-preview-tts to gemini-3.8-flash-tts is 2x input and 1.8x output. To gemini-3.8-flash-lite-tts it is 2x input and 1.2x output. gemini-3.1-flash-tts-preview or gemini-2.5-pro-preview-tts to gemini-3.8-flash-tts is the same input price with output 10% cheaper, and to gemini-3.8-flash-lite-tts output is 40% cheaper. Text input is a small share of a TTS bill, so the output change is the one to watch.
6
Free tier: the pricing page lists gemini-2.5-flash-preview-tts, gemini-3.1-flash-tts-preview and both 3.8 models as free of charge on the free tier. gemini-2.5-pro-preview-tts is listed as not available on the free tier.

Migration path

Recommended successor

gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts

Why this is the closest fit

They are the replacements Google names on all three deprecation rows. The speech generation guide says the two 3.8 models share the same API schema and prompting format, describes Flash as the pick for fidelity, acting and complex multi-speaker dialogue, and describes Flash-Lite as the high-throughput replacement for gemini-3.1-flash-tts-preview. Single-speaker and two-speaker output with prebuilt voices are both still supported.

What differs from the original

Three changes can silently break an existing pipeline. First, the text field is now a verbatim transcript: Google's guide says that unlike earlier preview models, where stage directions were embedded in plain text, 3.8 TTS separates turn-level directions into speech_metadata and keeps only point-in-time vocal events such as <short pause> as inline angle-bracket tags. A prompt like "Say cheerfully: welcome back" can now be spoken word for word. Second, audio format: the 3.1 preview and earlier TTS models returned headerless raw PCM (audio/l16) by default, while 3.8 returns WAV with a RIFF header on unary requests. If your code wraps the bytes in a WAV header itself, Google says to remove that wrapper. Streaming requests still return raw 24 kHz mono PCM. Third, multi-speaker: every turn in a two-speaker request must name its speaker in speech_metadata, and a single request supports up to 2 speakers using prebuilt voices. Custom designed or replicated voices are new in 3.8 and have to be synthesized one speaker turn at a time.

Alternatives by use case

You generate narration or voiceover for AI video clips on gemini-2.5-flash-preview-tts

Test gemini-3.8-flash-lite-tts first. It is 40% cheaper on output until December 31, 2026, and at 2027 prices it is the smaller increase (1.2x output versus 1.8x for 3.8 Flash)

You run gemini-3.1-flash-tts-preview or gemini-2.5-pro-preview-tts in production

Either replacement lowers or holds the per-second price even after January 1, 2027, so pick on voice quality: Flash for acted or multi-speaker scenes, Flash-Lite for bulk single-speaker reads

Your prompts put style instructions inside the text, such as "Read this in a whisper:"

Move those into speech_metadata or a designed voice before switching, and listen to the first outputs, because 3.8 reads the text field verbatim

Your code saves raw PCM and adds a WAV header with Python's wave module or ffmpeg

Remove the header step for unary calls, or request audio/l16 explicitly in response_format, so you do not end up with a double header

What this tool meant

gemini-2.5-flash-preview-tts, which Google's pricing page describes as optimized for price-performant, low-latency, controllable speech generation, will have run for about eighteen months as a preview, and stage directions written straight into the text were the whole steering method. The 3.8 replacements are generally available rather than preview, which is the usual trade: a stable model in exchange for a new prompt format and, from 2027, a higher price than the cheapest preview. As with OpenAI's tts-1 retirement, the named replacement covers capability, not your existing code, so the migration is a prompt and audio-handling change as much as a model string change.