ElevenLabs releases Eleven Realtime speech-to-speech models as Meta suspends Muse sharing

Voice AI led the day, with new real-time speech models from ElevenLabs and Voxtral 3 from Mistral. Meta paused Muse sharing after researchers found exposed conversations, and OpenAI moved deeper into pharma by acquiring Orchestra Bio.

1. Elevenlabs releases two Eleven Realtime speech-to-speech models

   ElevenLabs says its v1 Max variant scores 98.8 percent on Big Bench Audio with a 0.57-second response, claiming no ranked rival is both faster and more accurate.

2. Meta suspends Muse sharing feature after researchers expose shared conversations

   Claroc found sensitive personal and business data in a few hundred of 6,800 shared Muse links, prompting Meta to pause sharing and review the design.

3. Matrix-Game 4.0 renders playable game worlds at 480p and 60 FPS on one RTX 5090

   Pairing a lightweight game engine with a video-model renderer yields playable worlds on a single consumer GPU, with code and weights openly released.

4. Anthropic's Claude designs protein binders against scorpion neurotoxins

   New Anthropic research shows Claude designing protein binders aimed at scorpion neurotoxins, extending its use into biology.

5. Mistral launches Voxtral 3 speech models and a voice-agent SDK

   Mistral enters voice agents with priced transcription and text to speech plus an SDK, though it has not said whether weights are available.

6. NVIDIA releases Nemotron 3.5 Embed, Rerank and Parse retrieval models for agents

   NVIDIA's Nemotron 3.5 family adds embedding, reranking and document parsing models built for agentic retrieval.

7. OpenAI acquires Orchestra Bio to build drug development AI agents

   OpenAI will fold Orchestra Bio into its Rosalind life sciences platform to build agents that help pharma teams coordinate trials and spot problems.

8. CoEvo-RSI proposes harness and model co-evolution that lets Qwen3.5-9B compete with larger models

   Using only fine-tuning on its own rejection-sampled trajectories, Qwen3.5-9B reportedly beats DeepSeek-V4-Pro and GLM-5.3 on agent benchmarks.

Written by Claude (claude-opus-5-5) from the stories below.