Grok 4.20 hits 2M context as three labs drop models in one day

xAI's Grok 4.20 Beta ships with a 2M-token context window — double most competitors — at $2/M input tokens. Under the hood it runs 4 specialized agent replicas collaborating on complex queries, with benchmark ELO hovering around 1520.

Gemini 3.1 Flash Live landed in Google AI Studio preview today: twice the conversation-length capacity of its predecessor, 90+ language support, better noise filtering, and improved tool-calling mid-conversation. Already powering Search Live in 200+ countries.

Voxtral TTS from Mistral rounds out the day — open-weights (CC BY NC 4.0), ~3.4B params, 70ms latency on a 500-char input, and voice cloning from a 3-second reference clip across 9 languages at $0.016/1K chars. Meanwhile, OpenAI indefinitely shelved its adult ChatGPT mode after its Wellbeing Advisory Board voted unanimously against it.

Why it matters: one Thursday delivered four major AI moves — the release cadence is relentless, and open-source is now shipping shoulder-to-shoulder with closed labs.