Google's Gemini Robotics 2.5 hits 96% on live Mercedes-Benz lines
Google DeepMind's new Gemini Robotics 2.5 completed tasks 30% to 100% more successfully than previous baselines — and 20% to 30% faster — while Apptronik's Apollo 2 humanoid hit a 96% success rate doing automotive kitting on active Mercedes-Benz factory lines.
📊 Scorecard: our Bullish call on $GOOGL is scored from $287.38, the last close before it ran. Its first result comes after the next close.
The News
Google DeepMind introduced Gemini Robotics 2.5 on November 10, calling it its most capable vision-language-action model to date. A vision-language-action model is an AI system that can see its surroundings, understand instructions in plain language, and then physically act on them — the kind of brain a robot needs to do useful work in the real world instead of just answering questions on a screen. Across more than 200 tasks, the new model achieved success rates 30% to 100% higher than previous baselines, while completing tasks 20% to 30% faster.
The headline result came from a real factory, not a lab demo. In a proof-of-concept on active Mercedes-Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany, Apptronik's Apollo 2 humanoid robot performed automotive kitting — gathering the right parts into kits for assembly — at a 96% success rate. Kitting is the kind of repetitive, fiddly job that has long been hard to automate, because parts vary and the work needs both eyes and hands. A 96% success rate on a live line is the difference between a research project and something a factory manager can take seriously.
DeepMind also moved on safety, open-sourcing a benchmark called ASIMOV-Jailbreak. A jailbreak, in AI terms, is a trick that gets a model to ignore its safety rules; a benchmark is a standard test used to measure performance. On this test, Gemini 4 Argon scored 99.5% jailbreak resistance, compared with 82% for the prior baseline. The model itself is rolling out to partners only — including Agile Robots, Apptronik and Boston Dynamics — with no public API, which is the programming interface that would let outside developers build on it directly.
96% on a live Mercedes line — Gemini Robotics 2.5 is doing real factory work, not a lab demo.
Why It Matters
This matters because robots are crossing from demos into real work. A humanoid that can kit parts at 96% on an active car line is doing a job a factory actually needs done, on a line that cannot stop for a research experiment. The 20% to 30% speed gain matters just as much: a robot that is accurate but slow still costs a factory money. Faster and more reliable at the same time is the combination that turns a pilot project into a purchase order, and Mercedes-Benz is the kind of customer whose pilot, if it sticks, other manufacturers watch closely.
It also matters for Google. Gemini Robotics 2.5 is built on the same Gemini family as the company's flagship AI models, which means improvements in the core model flow straight into the robots. The partners-only rollout is a deliberate choice: DeepMind is putting the model in the hands of a small set of robotics companies first, learning from real deployments, rather than opening it up widely before it is ready. For Alphabet shareholders, the read-through is that Google's AI lead is showing up in physical work, not just chatbots — and the Mercedes-Benz link gives investors a second public company to watch, since the carmaker is the one running the live test.
The safety work matters too, and it is easy to overlook. A robot that can be tricked into ignoring its instructions is a liability in a factory full of people and expensive equipment. Publishing the ASIMOV-Jailbreak benchmark and the 99.5% figure is DeepMind saying it wants its safety claims tested in public. One caution: the 99.5% number comes from DeepMind's own benchmark, so it has not been independently replicated.
Analyst's Take
Our call is Bullish. A 96% success rate on a live Mercedes-Benz line, success rates up 30% to 100% across more than 200 tasks, and tasks done 20% to 30% faster add up to a real step forward, not a lab curiosity. The partners-only rollout keeps expectations honest — this is not a product you can buy tomorrow — but the direction is clear, and Google is the company turning its flagship AI into physical work first.
What to watch: whether the Mercedes-Benz pilot expands from proof-of-concept to a wider deployment, and whether other partners report similar results on their own lines. Alphabet closed Tuesday at $287.38, down 0.27%, and Mercedes-Benz closed at $11.56, down 1.78% — neither stock moved on this news yet, which means the market has not priced in a robotics win. If the pilots convert into orders, that changes.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is a vision-language-action model?
It is an AI that can see its surroundings, understand plain-language instructions, and then physically act — the kind of brain a robot needs to do useful work instead of just answering questions on a screen.
What is automotive kitting?
Kitting means gathering the right parts together into a kit for assembly — for example, collecting every bolt, bracket and connector a worker needs for one step of building a car. It is repetitive and fiddly, which has made it hard to automate.
What does a 96% success rate mean here?
It means Apollo 2 completed the kitting task correctly 96 times out of 100 on active Mercedes-Benz lines in Hungary and Germany, in a proof-of-concept run — a test on a real working line, not a lab.
What is ASIMOV-Jailbreak?
It is a safety test DeepMind open-sourced, meaning it published it for anyone to use. It measures how often an AI model can be tricked into ignoring its safety rules. Gemini 4 Argon scored 99.5% resistance, versus 82% for the earlier baseline.
Can I use Gemini Robotics 2.5?
Not yet. DeepMind is rolling it out to partners only — robotics companies including Apptronik, Agile Robots and Boston Dynamics — with no public API, the programming interface outside developers would use to build on it.
👁
Source: Google DeepMind