Best Text-to-Speech Models, September 2026: Ranked, Priced

By listener preference on the Artificial Analysis Speech Arena, read September 25, 2026, the best text-to-speech model is Cartesia's Sonic 3.6 at 1279 Elo. Second is Google's Gemini 3.8 Flash TTS, released this week, at 1265 and about a third of the price. On the other independent board, Voice Arena's US-English leaderboard, the cheaper Gemini 3.8 Flash-Lite TTS holds first place. The confidence intervals of the top four on Artificial Analysis overlap, so "best" is a tie broken by price and by the job you need done.

This page is the dated table: 92 models ranked by blind pairwise vote, every vendor's price normalised to dollars per million characters and per minute of audio, the licences of the 16 open-weight models, and a router by job. It is a snapshot. Both boards move weekly, and Google's launch prices double on January 1, 2027.

The Artificial Analysis text-to-speech leaderboard ranks models by Elo from blind pairwise votes in its Speech Arena, with a 95% confidence interval, across all prompt categories and both US and UK accents. The page carries no as-of date, so we date the snapshot to the day we parsed it. Price is Artificial Analysis' normalised figure in dollars per million characters at the endpoint it tracks. The table stops at rank 22, at 1130 Elo, tied with ranks 23 and 24; other vendors appear in the price table.

Read the intervals before the ranks. Sonic 3.6 at 1279 ± 17 and Gemini 3.8 Flash TTS at 1265 ± 17 overlap; so do Qwen-Audio-3.0-TTS Plus at 1259 ± 16 and Inworld's Realtime TTS-2 at 1246 ± 18. Simba 3.2 and Gemini 3.8 Flash-Lite TTS share 1239 exactly. Across the top six, price varies far more than Elo, a more than seven-fold spread: Simba 3.2 at $6.58 per million characters against Sonic 3.6 at $49. Artificial Analysis' own frontier note names three models on the quality-versus-price line: Sonic 3.6, Gemini 3.8 Flash TTS and Simba 3.2.

One normalisation deserves caution. Google bills its TTS models per audio token, not per character. Gemini 3.8 Flash TTS costs $9 per million output tokens and Gemini 3.1 Flash TTS costs $20, a 2.2× gap, yet Artificial Analysis' per-character figures for the two are $16.49 and $18.31, only 1.1× apart. We print the Artificial Analysis numbers for consistency and Google's own per-minute rate in the price table. Our post on the Gemini 3.8 Flash TTS launch works through the token arithmetic.

Google's launch post cited Voice Arena, whose US-English text-to-speech leaderboard uses Bradley-Terry Elo on prompts it says are built verbatim from real production calls. Its ranks come with a confidence-interval range, shown in brackets, and unlike Artificial Analysis it measures median time to first audio itself rather than repeating vendor claims. Rows where it has not yet measured latency say so.

The two boards agree on the cluster and disagree on the order. Sonic 3.6, Gemini 3.8 Flash TTS and Inworld's TTS-2 are top-four on both; Simba 3.2 is fifth on one and seventh on the other. The largest divergence at the top is Gemini 3.8 Flash-Lite TTS, first here by confidence-interval rank and sixth on Artificial Analysis. Its latency is not yet measured on Voice Arena, so the fastest measured models on this board are Simba 3.2 at 123 ms, Inworld's TTS 2 at 169 ms and Gradium at 236 ms. The Hugging Face TTS Arena, the third public board, is missing the September leaders, including Sonic 3.6 and Gemini 3.8, and was not used. Our guide to reading voice-AI benchmarks explains why arena votes and vendor demos diverge.

Vendors bill in four different units: per thousand characters, per million characters, per million audio tokens, and per UTF-8 byte. The table shows each vendor's own unit first, then our normalisation to dollars per million characters, then a per-minute figure. Where a vendor states its own per-minute rate we use it and say so. Where we derive it, we assume 750 to 1,000 characters per minute of speech, the range implied by Cartesia's and Hume's published ratios, so a $15 per million model costs about $0.011 to $0.015 a minute.

Three prices carry a date. Google's Gemini API pricing page lists the 3.8 TTS rates as promotional through December 31, 2026, with both input and output doubling on January 1, 2027. Deepgram's Flux TTS was free until September 12 and moved to $0.045 per thousand characters on September 13. And Alibaba's international price for Qwen-Audio-3.0-TTS could not be found on a primary page; the $27.59 figure is Artificial Analysis' endpoint price, and OpenRouter lists the same route at $20. Our Qwen-Audio-3.0-TTS post covers that model's arrival.

Sixteen of the 92 models publish weights. The licence column comes from each model's Hugging Face tag on September 25, 2026; the commercial column is Artificial Analysis' classification, which is not a legal opinion and in one case conflicts with the tag. The pattern is clear: the open models that score well are non-commercial, and the commercially usable ones score around 1063, about 200 Elo below the leaders.

Kokoro 82M is the practical answer for a self-hosted or cost-floor deployment: Apache 2.0, 11.7 million Hugging Face downloads, and $0.62 to $0.65 per million characters hosted, which Artificial Analysis names as the cheapest endpoint on its board. Voxtral TTS is the strongest open model by vote from a major lab, but its CC BY-NC licence rules out commercial use without a separate agreement; our Voxtral comparison has the detail.

The right model depends on what the audio is for. The router below is our reading of the two boards and the price table; every latency figure in it is Voice Arena's measurement, and every vendor claim is labelled as one.

For a realtime conversational agent the choice is different again, because speech-to-speech models skip the text step. Our voice agent infrastructure reference covers that stack, and our content engine practice builds narration and dubbing pipelines on the models above.

Two rules apply to anyone shipping synthetic audio at scale. In the European Union, Article 50 of the AI Act has applied since August 2, 2026: providers of systems that generate synthetic audio must mark outputs in a machine-readable, detectable form, and deployers of deepfakes must disclose them at first exposure. The Commission's Article 50 FAQ, updated July 24, 2026, says that disclosure cannot rely only on embedded marks and must be perceivable, for example through audible labels. Fines run to €15 million or 3% of worldwide turnover. The Digital Omnibus gives systems placed on the market before August 2, 2026 until December 2, 2026 to add the marking.

In the United States, the FCC's declaratory ruling of February 8, 2024 classed AI-generated voices as artificial or prerecorded voices under the Telephone Consumer Protection Act, so an outbound call using any model in this table needs prior express consent.