🤖 OpenAI ships GPT-Live, its full-duplex voice model 🤖
OpenAI begins rollout of a voice model that listens, thinks, and speaks simultaneously.

OpenAI has released GPT-Live, a next-generation voice model built on a full-duplex architecture that processes inputs continuously — meaning it can listen and speak at the same time, handling natural conversational cues like interruptions and pauses. The model delegates complex tasks to GPT-5.5 for background processing and surfaces visual cards alongside spoken answers. In internal head-to-head tests, users preferred GPT-Live over the previous Advanced Voice Mode roughly 75% of the time, while expert science scores nearly doubled. Real-time voice interaction is now a core competitive differentiator for OpenAI. The model also supports real-time translation and is expected to roll out across all ChatGPT tiers in the coming weeks.

🚀 Grok 4.5 launches free inside Cursor 🚀
Aggressive pricing and Cursor integration put Grok 4.5 against Opus-class rivals.

SpaceXAI launched Grok 4.5, the first model co-developed with Cursor following xAI's $60 billion acquisition, targeting coding, agentic tasks, and knowledge work. Running at 80 tokens per second, the model claims a 4× efficiency gain over competitors and is priced at just $2/$6 per million input/output tokens — compared to $5/$25 for Opus 4.8. This aggressive pricing pressures the entire enterprise AI market. Usage is temporarily free inside Cursor and the new Grok Build environment. Elon Musk previewed an even larger model arriving next month, positioning this release as the opening move in a broader push to combine Cursor's coding-tool dominance with xAI's model capabilities.

💻 Cognition's SWE-1.7 matches GPT-5.5 at lower cost 💻
Near-frontier coding intelligence arrives at a fraction of major-lab pricing.

Cognition released SWE-1.7, an in-house coding model that powers its Devin autonomous coding agent, built on top of China's open-source Kimi K2.7 foundation. The model reaches near-frontier intelligence levels comparable to GPT-5.5 and Opus-class systems — but at substantially lower cost, proving smaller labs can compete using open-source foundations. Broad improvements across Cognition's reinforcement learning pipeline, including better infrastructure, more stable training, and novel techniques for long-horizon coding tasks, drove the gains. SWE-1.7 is available through Devin on web, desktop, and CLI, and could significantly expand access to advanced AI-powered software engineering for development teams.

INTERESTING TO KNOW

🎨 ByteDance rolls out Seedream 5.0 Pro image model 🎨

ByteDance has opened access to Seedream 5.0 Pro, a multimodal image model that goes beyond generation to "understand design," offering precision editing with layer separation, multilingual support across 10+ languages, and production-ready creative workflows. Positioned as a direct competitor to OpenAI's Images 2.0, its design-aware approach challenges existing dominance in creative AI tools and could appeal to professional teams. The release arrives ahead of ByteDance's anticipated Seedance 2.5 video model.

🤖 Mistral ships Robostral Navigate for robot vision 🤖

Mistral has released Robostral Navigate, a compact 8-billion-parameter model enabling robots to autonomously navigate physical environments using only a single RGB camera. This integration of foundation models into physical-world applications marks Mistral's first major foray into embodied AI, expanding European AI presence in a robotics space largely dominated by US and Chinese players. By requiring minimal sensor hardware, the approach lowers the barrier for robotics developers significantly.