Gemini Robotics ER 2 is live in the Gemini API and AI Studio
Google's new embodied-reasoning model reads video and coordinates robots. Also: 90%-off caching for GPT-5.6 on Bedrock, and GCC bars AI-written code.
Copy markdown
A robot-brain model you can call from the Gemini API
Google DeepMind's Gemini Robotics ER 2 shipped July 30 in the Gemini API and AI Studio (preview). It reads continuous video to plan and track multi-step tasks — spotting key moments at 91.3% accuracy with 0.96s latency — and can hand work between wheeled, humanoid, and arm robots over the Live API. It's the same standard Gemini endpoint, so builders can wire spatial reasoning and video understanding into agents today.
GPT-5.6 caching lands on Bedrock: 90% off repeat reads
AWS turned on explicit prompt caching for GPT-5.6 Sol, Terra, and Luna on Amazon Bedrock. Mark up to 4 prompt_cache_breakpoints (min 1,024 tokens each); cached reads cost 90% less and stay warm for at least 30 minutes, with cache writes at 1.25x. For long agent loops that resend a big fixed system prompt, that erases most of your input bill.
GCC won't take your AI-written compiler code
The GCC steering committee will now decline any legally significant contribution generated by — or derived from — an LLM, with one carve-out: AI-written test cases are fine if clearly marked. The policy aligns with GNU practice and gets revisited in early 2027. If you contribute to GNU projects, check each one's line before opening an AI-assisted PR.
Elsewhere: Zed 1.13.1 adds thinking and shareable dev containers
Zed's editor moved to Mistral Medium 3.5 with a thinking mode (plus Mistral Small 4), and its dev containers can now be reused by VS Code and the CLI. Docker Compose resumes now start every configured service, and the artificial 17-second git-timeout floor is gone.