Gemini Robotics 2.5: 96% success on a Mercedes-Benz production line
Direct answer
Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, its most capable vision-language-action model for robots. In a proof of concept, Apptronik Apollo 2 humanoid robots sorted metal parts on live Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, with a 96% success rate, according to DeepMind. The job was kitting: taking metal brackets out of bulk bins and placing them into trays in the exact order each vehicle on the line needs. The figure is DeepMind's, with no trial count published, and the test is a proof of concept, not a plant-wide rollout.
What Gemini Robotics 2.5 is
- A model that sees, understands and acts: Gemini Robotics 2.5 is a vision-language-action model: it takes camera images and an instruction, and outputs robot motions. DeepMind evaluated it across more than 200 distinct tasks, from serving food with tongs to pipetting liquid and tying a bow on a gift box.
- Improvements measured against the previous version: according to DeepMind, success rates are 30% to 100% higher than the July-era Gemini Robotics 2 baseline, and completion speeds are 20% to 30% faster. These are the company's figures, from its own tests, with no independent verification.
- The 96% has context: Ars Technica notes that a human doing a repetitive task like this gets it right almost every time, so 96% is still short for a manufacturer. DeepMind director Kanishka Rao told Ars Technica that closing the remaining 4% is a steep road. The current model runs on cloud servers, which adds latency, and the demonstration videos are sped up.
- The figure DeepMind does not publish: the announcement does not say how many trials the 96% covers, how DeepMind defines a success, or how often a human stepped in. A 96% can be 24 successes in 25 attempts or 960 in 1,000, and the confidence range changes completely. The figure describes reported performance under the trial conditions.
The other news: a public test for tricking robots
- ASIMOV-Jailbreak is an open test: alongside the model, DeepMind published ASIMOV-Jailbreak, an open-source dataset of prompts designed to convince a robot to do something unsafe. It is the industry's first shared yardstick for measuring robot jailbreak resistance.
- Gemini 4 Argon resists 99.5% of attempts: according to DeepMind, Gemini 4 Argon resists 99.5% of jailbreak attempts, against 82% for the previous baseline. The figure is DeepMind's, from its own test set.
- Robot safety is governance, not an accessory: a robot that executes instructions in the physical world needs the same controls as an agent that executes actions in your systems: permissions, limits and a record of every action. The open test set gives the industry a common way to measure that control.
- Rollout goes to partners, not the general public: DeepMind says Gemini Robotics 2.5 is rolling out to robotics partners and names Agile, Apptronik and Mercedes-Benz among the teams that supported the release. There is no general-availability date or public pricing.
What to do in your business this week
- Read the 96% as a pilot signal: the figure says the technology works on a real line, on one concrete task, in two plants. The figure does not say humanoids are ready for your plant. The question for your business is which repetitive task in your operation resembles kitting: ordered parts, fixed sequence, controlled environment.
- Demand the denominator before deciding: when a provider shows you a success rate, ask for the trial count, the definition of success and the human-intervention frequency. Without a denominator, the rate is a headline, not a metric.
- Start with one task, not one plant: the Mercedes-Benz test is a single task on a live line. The entry path for your company is a one-task pilot, with your own metrics, before talking about rollout.
- Put physical safety in your AI governance: if a robot will act in your operation, the plan includes permissions, action limits and logging. The ASIMOV-Jailbreak test set is a public reference for measuring that resistance.
Note: this article summarizes Google DeepMind's announcement of November 10, 2026 and coverage by Ars Technica, The Decoder, TechTimes, Modern Robotics and Industrial News. The success and improvement figures are DeepMind's, from its own tests, with no independent verification. No Mercedes-Benz confirmation of the 96% figure was located.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- Google opened Gemini 4 Argon with launch pricing: $2 per million tokens until Nov 30
- OpenAI launched Daybreak Red Team and Defense Factory: agents that attack your code and then repair it
- Anthropic opens the beta of self-hosted Claude Managed Agents: the agent runs in your cloud
Frequently asked questions
What is Gemini Robotics 2.5?
It is Google DeepMind's most capable vision-language-action model, introduced on November 10, 2026. It takes camera images and an instruction, and outputs robot motions. It was evaluated across more than 200 distinct manipulation tasks.
What did it achieve at Mercedes-Benz?
In a proof of concept, Apptronik Apollo 2 humanoid robots sorted metal brackets on live Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany, with 96% success, according to DeepMind. The job was kitting: placing parts into trays in the order each vehicle needs.
Does 96% mean the robots are ready for the factory?
The figure is DeepMind's, with no trial count published, and the test is a proof of concept. Ars Technica notes that a human gets this task right almost every time, and DeepMind's director called the road to closing the remaining 4% steep.
What is ASIMOV-Jailbreak?
It is an open-source test set published by DeepMind alongside the model, with prompts designed to convince a robot to do something unsafe. DeepMind says Gemini 4 Argon resists 99.5% of attempts, against 82% for the previous baseline.
