Playbooks · Coding
Gemini Robotics 2.5: 96% packing, partners only
Written November 11, 2026
DeepMind calls Gemini Robotics 2.5 its most capable vision-language-action model to date. It was tested across more than 200 tasks and packed car parts on a working factory line with a 96% success rate. You cannot use it: the rollout is partners only, with no public API, price, or release date.
The problem
A robot that works in a lab demo and a robot that works on a factory floor are two different things. Factory floors are unstructured: bins are messy, parts vary, and people walk past. A headline success rate from a vendor pilot is only useful if you know what was measured and what the announcement does not say.
What 2.5 does
DeepMind's announcement describes a model that takes in what a robot sees, reasons about the task, and acts. It works across different robot bodies, including Apptronik's Apollo 2 humanoid, with a five-finger hand, and the Franka Duo platform, with two-fingered grippers. Across more than 200 distinct tasks — serving food with tongs, tying a bow on a gift box, pipetting liquid — DeepMind reports success rates 30% to 100% higher than previous baselines, depending on the type of task, while completing tasks 20% to 30% faster. Gemini 4 Argon helped build it: Argon generated simulated practice runs from real camera scenes and helped researchers work out why failures happened.
The factory pilot
The strongest claim in the announcement comes from a proof of concept on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 robots did automotive kitting: sorting individual metal brackets from bulk bins into sequencing containers in the exact order each vehicle on the line needs. An embodied reasoning model kept track of the workflow while Gemini Robotics 2.5 controlled the robot's hands and arms. DeepMind reports a 96% success rate.
What the 96% does not tell you
Read the number with its gaps attached. DeepMind does not publish how many trials the 96% covers, how it defines success for a pick-and-place cycle, or how often a person had to step in. The figure is DeepMind's own measurement of its own pilot, not an independent test. It shows the model can do repetitive kitting work on a real line; it does not show how it behaves on a task it has never seen, or what happens when the line changes.
Safety work in the same release
DeepMind also reports safety results. Inference-time reasoning guidance that steers around physical hazards improved safe task completion around obstacles by 33% and around nearby people by 43%. The team released ASIMOV-Jailbreak, an open benchmark of prompts that disguise unsafe requests across dozens of scenes. On it, Gemini 4 Argon scores 99.5% jailbreak resistance, against 82% for the previous baseline, Gemini Robotics ER 2.
What to do
Nothing, for now, unless you are one of the partners. The rollout covers Agile Robots, Apptronik, Boston Dynamics, and more than 100 trusted testers. There is no public API, no price, and no general-availability date. If you build with robots, the useful part of this release is the benchmark and the pilot shape. ASIMOV-Jailbreak is open and can be run against your own setup. The kitting pilot is a concrete example of the kind of repetitive, physically taxing task these models are aimed at first — work that frees technicians for complex build tasks rather than replacing them.
Limits
- Vendor-reported, partners-only. The 96% figure has no published trial count, success definition, or human-intervention count, and no independent measurement exists.
- The 30–100% improvement range is against DeepMind's own previous baselines, and the range is wide because it varies by task type.
- Nothing here was run by us: no robot, no benchmark, and no pilot was tested for this page. The claims are DeepMind's, as published.
Related playbooks
- The intelligence-explosion argument, and its evidence
- Gemini 4 Argon doubles in price on December 1
- Gemini 4 Carbon is here: $2 in, $10 out
- Run Gemini 4 Hydrogen from the Copilot picker
Sources
Back to: Coding playbooks hub · Also: Claude designs three electronic prototypes · Claude Meetings, Sheets and Whiteboards land