AI news
Language
AirCaps A5S Async v1 tops independent benchmark for who-said-what transcription
It ranked first of 15 systems on VoiceArena's Diarization-Bench, covers 9 languages and costs $0.18 per audio hour. Meeting notes, interview transcripts and subtitles depend on getting the speaker right, not only the words.
Models
Anthropic releases Claude Fable 5.2 and Mythos 5.2 at lower prices
Up to 30% faster and down to $8 per million input tokens and $40 output, with about 55% fewer unnecessary safety blocks. Anthropic still advises Opus 5.5 for everyday work. Follows the 2 Sep Fable 5.1 release.
Models
Google DeepMind unveils Gemini Robotics 2.5, rolling out to robotics partners
An Apptronik humanoid running it hit a 96% success rate sorting parts on Mercedes-Benz lines in Hungary and Germany. DeepMind also open-sourced ASIMOV-Jailbreak, a benchmark for disguised unsafe robot instructions.
Models
Google makes Gemini 4 Carbon available to developers, businesses and Ultra users
Follows the 28 Oct preview for cyber defenders only. API pricing is $2 and $10 per million tokens until 30 Nov, then $4 and $20, and a /swarm preview in Antigravity runs teams of agents from one prompt.
Models
Meta launches Muse Ember, a low-cost model for high-volume agent work
At $0.10 per million input tokens with a 1M-token context, it matches Claude Haiku 5.5 on Artificial Analysis at about a third of the cost. Built to run repeated subtasks while a larger model plans.
Communication
Nuance Labs launches Emma, a video AI persona that listens and talks at once
Built on its Nuplex full-duplex model, Emma nods and reacts mid-sentence, and a beta API lets developers brief their own personas. English only, with watermarks and no cloning, and available in the UK and EU.
Language
Zyphra releases ZONOS3-Mini, an open voice-cloning TTS model that runs on phones
Apache 2.0 weights clone a voice and steer accent, emotion and pace word by word, in real time in a phone browser. Cheap on-device cloned speech helps dubbing and voice teams and sharpens consent questions.