Gemini Robotics 2.5 explained: 30-100% better, partners only

By LLMTest Lab · Nov 11, 2026 · 3 min read releaseroboticsgoogle
On this page

On this page

  1. What is new
  2. What DeepMind reports
  3. Safety claims
  4. Who should care

Not tested by us yet. Gemini Robotics 2.5 drives robot bodies, and our harness only calls text and image models through an API, so we cannot run it. DeepMind says it is "rolling out to our robotics partners", with no public API or price. Everything below comes from DeepMind's launch post. We will update this post when access opens or independent results appear.

Gemini Robotics 2.5 is Google DeepMind's most capable vision-language-action (VLA) model, announced on November 10, 2026. DeepMind reports 30% to 100% higher task success than Gemini Robotics 2, depending on task type, and tasks finished 20% to 30% faster. It is for robotics teams with partner access. If you are picking chat or coding models, it does not change your shortlist.

Gemini Robotics 2.5
Maker Google DeepMind
Announced 2026-11-10
Type Vision-language-action model for robots
Predecessor Gemini Robotics 2
Evaluation Over 200 distinct tasks (DeepMind-run)
Robots shown Apptronik Apollo 2 humanoid, Franka Duo
Availability Rolling out to robotics partners
Public API and price None announced

What is new

A VLA model takes camera input and an instruction and outputs robot actions. Version 2.5 is built to learn more skills, and to learn them faster, than version 2. DeepMind says it works natively across many robot types and shows two: the Apollo 2 humanoid with a five-finger, 22 degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers.

The post also leans on the wider Gemini family. DeepMind says Gemini 4 Argon served as a simulation engine, generating practice runs and digital twins to speed up development. If you want the language-model side of that family, our Gemini 4 Carbon review covers its sibling.

What DeepMind reports

All of these are DeepMind's own numbers, and nobody independent has reproduced them.

Task Gemini Robotics 2 Gemini Robotics 2.5
Close lid of black toolbox 0% 93%
Unhang mug to table 29% 75%
Sweep trash into dustpan with brush 0% 52%
Pack rice cooker 0% 40%
Pack pastries with tongs 5% 100%
Untie ribbon bows 0% 70%
Hang a cardigan 0% 62%
Fold pants 0% 40%

Read the table with care. Several baselines are 0%, so "30% to 100% higher" describes a spread across task types, and a jump from 0% to 40% on pants folding still means the robot fails more often than it succeeds. The tasks were chosen and scored by the lab that built the model.

The one deployment-flavoured number is a proof of concept on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. There, Apollo 2 picked and packed automotive parts with a 96% success rate. That is a pilot, not a production commitment, and DeepMind gives no throughput or cost figures.

Safety claims

DeepMind says inference-time reasoning guidance improved safe task completion by 33% around obstacles and 43% around nearby people. It also introduces ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts across dozens of scenes, and reports that Gemini 4 Argon reaches 99.5% jailbreak resistance against 82% for the previous baseline, Gemini Robotics ER 2.

The benchmark being open is the useful part. Outside teams can run it and publish numbers that do not depend on DeepMind's setup. The 99.5% figure is for Argon, the reasoning model, not for the action model itself, so do not read it as a statement about the robot policy.

Who should care

Teams already in Google's robotics partner program are the audience. Everyone else cannot use it yet. Compared with Nvidia's open GR00T 2, the trade is plain: DeepMind reports stronger per-task gains against its own predecessor but offers no weights, no license and no public endpoint, while Nvidia describes its model as open, though it has not yet said where to download it.

Wait if you need something you can budget, benchmark or ship. With no price, no access date and no independent evaluation, the numbers above are marketing-grade evidence until someone else repeats them.

We will update this post once Gemini Robotics 2.5 is generally available or third-party results appear.

Related articles

Gemini 4 Carbon review: beats Argon 5-0, ties Fable 5.2 at 1/5 cost
We ran Gemini 4 Carbon against Argon and Claude Fable 5.2 on 8 hard engineering prompts. It won 5-0 and 2-2 at $2/$10 per million tokens, doubling after Nov 30.
GR00T 2 explained: Nvidia's open world action model for robots
GR00T 2 is Nvidia's open world action model for humanoid robots. What Nvidia reports, what is still missing, and who should look at it first.