Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS is Google’s text-to-speech model for directed, expressive speech, announced on Google’s blog on 23 September 2026; the Gemini API changelog lists its general availability a day earlier. It turns text into audio and is aimed at production work: a voice can be designed from a written description, delivery can be directed line by line, two-speaker scenes are generated natively and long-form audio is supported. Google says more than 2,000 ready-made voices are available, and with a consent check a voice can be replicated from a 30-second sample — though replication is not offered in the UK, the EEA, Switzerland, India, Illinois or Texas. The model page lists 130 languages, an 8,192-token input limit and a 16,384-token output limit, with batch, caching and flex or priority inference, but no function calling, structured output or grounding. Output carries Google’s SynthID watermark. The price is the detail to plan around: on the paid tier it is $0.50 per million input tokens, $9.00 per million audio output tokens (one second of audio is 25 tokens) and $0.125 per million cached input tokens through 31 December 2026, and every one of those rates doubles on 1 January 2027 — $1.00, $18.00 and $0.25. Google’s pricing page also marks free-tier use as used to improve its products, and paid-tier use as not. On Artificial Analysis’s independent Provider Voice Arena it ranked second, with an Elo rating of 1,260, when read on 24 September 2026; arena ratings move as votes arrive. Google’s own claims — first place on Hume AI’s Voice Design Benchmark and leading positions in several languages — are its own reading and are not stored as scores.