Google DeepMind introduces Gemini Robotics ER 2 for multi-step task planning
Saturday, August 1, 2026 · what moved in the AI release cycle in those 24 hours.
One click creates your free account (or signs you into an existing one) and turns on alerts for the Daily Briefing. There's no password to set, and you can unsubscribe any time.
Google DeepMind unveiled Gemini Robotics ER 2, a model it says can plan multi-step tasks, orchestrate lower-level robotics tools, and use continuous video to track progress and adapt to failures. OpenAI, meanwhile, offered a technical explanation for GPT-5.6 Sol's improved ARC-AGI-3 showing, attributing a jump from 13.3% to 38.3% on the public set to retained reasoning and compaction in its Responses API harness. Elsewhere, benchmark scores trickled in for Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Kimi K3.
An AI summary of the day’s tracked items. It may contain errors; every item below links to its source.
A new embodied-reasoning model from a frontier lab signals Google's continued push into robotics-oriented AI, with claimed multi-step planning and video-based error recovery.
OpenAI's explanation that harness changes nearly tripled GPT-5.6 Sol's ARC-AGI-3 score highlights how much evaluation setup, not just model weights, drives reported capability.
Google says the model can plan multi-step tasks, orchestrate lower-level robotics tools, and use continuous video to track progress and adapt to failures.