Google DeepMind has introduced Gemini Robotics 2.5, which it calls “our most capable vision-language-action (VLA) model to date”. In its announcement on 10 November, DeepMind says a proof-of-concept on Mercedes-Benz production lines saw Apptronik’s Apollo 2 humanoid “pick and pack automotive parts with a 96% success rate”.
The same post reports a safety result for Gemini 4 Argon, Google’s frontier model: “Argon achieves a 99.5% jailbreak resistance”, against “the 82% rate of our previous baseline, Gemini Robotics ER 2”. That means the share of disguised unsafe prompts that got through fell from 18% to 0.5%, 36 times fewer, by our arithmetic. Gemini Robotics 2.5 “is rolling out to our robotics partners”; the post does not mention an API or a price.
Business Pill · Physical AI Explained
A short explainer on physical AI, models that act through robot bodies. It teaches the general idea only and does not discuss Gemini Robotics 2.5.
The key insight:Google DeepMind says Gemini Robotics 2.5 helped Apptronik’s Apollo 2 humanoid pick and pack automotive parts at a 96% success rate in a Mercedes-Benz proof-of-concept. In the same post, Gemini 4 Argon resisted 99.5% of disguised unsafe prompts against 82% for Gemini Robotics ER 2: the share that got through fell from 18% to 0.5%, 36 times fewer, by our arithmetic.
DeepMind says it tested the model “across over 200 distinct tasks”, such as “serving food with tongs, tying a bow on a gift box, and pipetting a liquid”.
Its headline performance figure: “Compared to previous baselines, Gemini Robotics 2.5 achieves a 30% to 100% higher success rate depending on the type of task, while completing tasks 20% to 30% faster”. The post does not give per-task success rates or name the baseline model for that comparison.
The model works with different robot types, DeepMind says, including Apollo 2 with the “five-finger, 22 degree-of-freedom SharpaWave hand” and the “Franka Duo platform with two-fingered parallel grippers”. Earlier this year, the post says, DeepMind “brought whole-body intelligence to robots of many types, including humanoids”; its related-posts list dates Gemini Robotics 2 to July 2026.
Jailbreak resistance on DeepMind’s ASIMOV-Jailbreak benchmark as stated in its 10 November 2026 post: Gemini Robotics ER 2 82%, Gemini 4 Argon 99.5%. The share that got through is our arithmetic (100% less resistance).
What Was Posted on X
Gemini Robotics 2.5 is here. ๐ค
Our latest physical AI is able to learn hundreds of everyday skills – from folding laundry and serving pastries to autonomously sorting car parts on the factory floor.
Google DeepMind (@GoogleDeepMind) posted on 10 November 2026 at 15:00 UTC: “Gemini Robotics 2.5 is here.” The post describes skills from folding laundry and serving pastries to sorting car parts on the factory floor. The figures in this piece come from DeepMind’s blog post.
The Mercedes-Benz Pilot
DeepMind says it partnered with Apptronik and Mercedes-Benz to use Apollo 2 humanoids “on active automotive production lines in Kecskemรฉt, Hungary, and Sindelfingen, Germany”. The task was kitting: “sorting individual parts from bulk bins into organized trays in the exact order needed for each vehicle on the line”.
Two models split the work. The embodied reasoning model, Gemini Robotics ER, keeps track of the workflow, and the Gemini Robotics 2.5 actions model controls the hands and arms, the post says. The robot “reaches into supply bins, picks up metal automotive brackets, and places them into sequencing containers”.
Apptronik’s co-founder and chief executive, Jeff Cardenas, is quoted in the post: “we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off”. The post does not say how many robots, shifts or parts the pilot covered.
Safety: Steering and Jailbreaks
DeepMind says Gemini Robotics 2.5 uses “inference-time reasoning guidance” to steer around physical hazards. “In our evaluations, this capability boosted safe task completion around obstacles by 33% and around nearby people by 43%”, the post says.
It is also releasing ASIMOV-Jailbreak, “a new open-source benchmark containing hundreds of prompts that disguise unsafe requests across dozens of scenes”. The 99.5% and 82% figures come from that test. The 99.5% is Argon’s, not a score for the Gemini Robotics 2.5 actions model.
Built With Gemini 4 Argon
DeepMind credits Argon for much of the gain: “in large part thanks to Gemini 4 Argon’s ability to accelerate robotics research”. Argon “can generate thousands of varied, simulated practice runs from a real-world camera scene and build a realistic digital twin”, the post says.
Google introduced Gemini 4 Argon for Robotics on 13 October, when it said Argon scores 87% on the Embodied Reasoning (ERQA) benchmark. The new post names Agile, Apptronik and Mercedes-Benz as partners on the release.
As we read the post, the release pairs two kinds of evidence: a factory number from one proof-of-concept and benchmark numbers from DeepMind’s own evaluations, and neither has been independently replicated.
As we read the figures, the jailbreak result is the most precisely stated: two percentages on a named, open-source benchmark, while the success-rate gain is a 30% to 100% range against baselines the post does not name.
Jeff Cardenas, co-founder and CEO of Apptronik, in Google DeepMind’s post, 10 November 2026
“This is the foundation for bringing humanoid robots to industrial and commercial environments at scale”
Three Implications
FOR FACTORIES The pilot task was kitting at Mercedes-Benz plants in Kecskemét and Sindelfingen, at a 96% success rate, DeepMind says; the post gives no pilot size or duration.
FOR SAFETY Argon resisted 99.5% of ASIMOV-Jailbreak prompts against 82% for Gemini Robotics ER 2; 0.5% against 18% got through, by our arithmetic.
FOR ACCESS Gemini Robotics 2.5 is rolling out to robotics partners, the post says; it names no API or price.
The framework describes “The FRED Test โ the diagnostic for whether a workflow survives agentic automation. Frequency, Repeatability, Error tolerance, Decomposability.”
As we read it, the Mercedes-Benz task fits the first, second and fourth tests by DeepMind’s own description: kitting is repetitive work, sorting parts into trays in a set order. The 96% success rate is the error-tolerance question the post leaves open: it does not say what a 4% miss rate means on the line. The framework was published in June and does not discuss Gemini Robotics.
What Is Not Established
The figures are DeepMind’s own evaluations and have not been independently replicated. The 96% is from one proof-of-concept, and the post does not give its size or duration. The 30% to 100% gain is a range across task types against unnamed “previous baselines”. The post gives no availability date beyond robotics partners, and no pricing. The 36 times figure is our arithmetic from DeepMind’s two percentages.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Gemini Robotics 2.5 comes with a factory number, 96% success on Mercedes-Benz kitting, and a safety number, Argon’s 99.5% jailbreak resistance against 82% for Gemini Robotics ER 2. In Google’s test, 0.5% of disguised unsafe prompts got through, down from 18%.
A note on sourcing. Figures and quotations are from Google DeepMind’s announcement of 10 November 2026 and Google’s Gemini 4 Argon for Robotics post of 13 October 2026; they are the company’s own evaluations. The share of jailbreaks that got through and the 36 times ratio are our arithmetic. Google was not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
