Model releases · Google DeepMind ·
DeepMind’s Gemini Robotics 2.5 hits 96% success in a Mercedes-Benz factory pilot
A new vision-language-action model, tested on 200 real-world tasks and on humanoid robots doing automotive kitting.
What happened
Gemini Robotics 2.5 is DeepMind’s most capable vision-language-action model so far. Across 200 real-world tasks — serving food with tongs, pipetting, gift-wrapping — DeepMind reports 30–100% higher success rates than earlier baselines and 20–30% faster completion.
In a proof of concept with Mercedes-Benz and Apptronik, Apollo 2 humanoids running the model reached 96% success on automotive kitting in factories in Hungary and Germany. DeepMind described closing the last 4% as a “steep road”.
DeepMind also released ASIMOV-Jailbreak, an open benchmark for testing whether robot models can be manipulated into unsafe physical actions.
What it means for you
Robotics models are moving from demos to factory pilots; the open safety benchmark is the piece researchers can use today.
Sources
Written by TheEvents from public reporting and the companies’ own announcements. Facts can change as stories develop — check the sources for the latest.
Related
OpenAI’s GPT-6 Cyber built working root exploits for 948 Linux bugs
A security-specialised model rated “Critical” under OpenAI’s Preparedness Framework, with access limited to vetted security teams.
Gemini 4 Carbon opens to all paying developers — intro price ends 30 Nov
Google’s most capable model is now in the Gemini API and for AI Ultra subscribers, at $2/$10 per million tokens until 30 November.
Anthropic releases Claude Fable 5.2 and Mythos 5.2
Anthropic’s new model generation targets advanced and long-running tasks, landing the same day as Claude’s new Meetings and Sheets features.
Gemini 4 Argon launches as Google’s new everyday frontier model
Gemini 4 Argon arrives in AI Studio, Antigravity and Android Studio at $2/$10 per million tokens, alongside a new Gemini API CLI.