Skip to content
IFAI

Model rankings

912 models · 5 boards · 2 sources

Text-to-speech arena · Anonymous head-to-head votes · Updated January 9, 2026

Best voice
Inworld TTS 1 MaxInworld1,170Score
Best open model
Kokoro 82M v1.0Kokoro1,060Score
Best from China
Speech 2.6 HDMiniMax1,153Score
Top three by Score
1Inworld TTS 1 MaxInworld1,170
2Speech 2.6 HDMiniMax1,153
3Speech 2.6 TurboMiniMax1,148
Top three open models
1Kokoro 82M v1.0Kokoro · #15 overall1,060
2Fish Speech 1.5Fish Audio · #30 overall1,030
3Maya1Maya Research · #31 overall1,028
This week's changes
Doubao-Seed-1.8Intelligence · #23New
Magpie-Multilingual 357MSpeech · #35New
Mi:dm K 2.5 ProIntelligence · #66New
RankChangeModelScore95% interval
1—TTS-1ClosedOpenAI1,107—
2—Polly GenerativeClosedAmazon1,061—
3—TTS-1 HDClosedOpenAI1,059—
4—Azure NeuralClosedMicrosoft Azure1,051—
5—Polly Long-FormClosedAmazon1,048—
6—StudioClosedGoogle1,047—
7▼ 1GPT-4o mini TTSClosedOpenAI1,047—
8—JourneyClosedGoogle1,038—
9—Chirp 3: HDClosedGoogle1,024—
10NewMagpie-Multilingual 357MOpenNVIDIA1,014—
11▼ 1Magpie MultilingualClosedNVIDIA1,000—
12▼ 1VibeVoice 1.5BClosedMicrosoft Azure991—
13▼ 1VibeVoice 7BOpenMicrosoft Azure969—
14▼ 1Polly NeuralClosedAmazon937—
15▼ 1WaveNetClosedGoogle906—
16▼ 1StandardClosedGoogle877—
17▼ 1Neural2ClosedGoogle874—
18▼ 1Polly StandardClosedAmazon849—

Intelligence over time

Each maker's best model at the time on the Artificial Analysis Intelligence Index, by the month it was released.

Gained most · Anthropic +57.2Highest now · Anthropic 64.6Gap between highest and lowest 18.1 ptsTime range Nov 2025 – Nov 2026
64.6Anthropic60.0Google56.9OpenAI50.1Meta46.5Kimi

Arena data comes from the leaderboard dataset it publishes on Hugging Face (CC BY 4.0). Only each board's overall leaderboard is used; scores and ranks are unchanged. dataset · CC BY 4.0

Artificial Analysis data comes from its API (artificialanalysis.ai). Scores are unchanged; ranks follow the scores. artificialanalysis.ai

Open or closed follows the license Arena gives; prices and Artificial Analysis intelligence show only where model names match, otherwise “—”.