Google has unveiled Gemini 4 Argon, a new model that it says reduced the resource requirements of a quantum computing operation by more than 100 times, while helping researchers develop new mathematical solutions for nuclear fusion and improve real-world robot performance.
In one demonstration, Argon generated a quantum algorithm component with more than 100 times lower spacetime cost than a previously published method. The company said the implementation was formally verified using Lean, a system for checking mathematical proofs.
The model also helped researchers identify new solutions to a decades-old mathematical problem related to stellarator fusion reactors. In separate robotics tests, training data generated through Argon’s simulations increased a robot’s real-world task success rate from 56% to 75%.
Google announced the developments on October 13 as part of the Gemini 4 Argon launch. In another test, the model worked unattended for 11 hours on Google’s existing software codebase, producing changes that reduced session preparation overhead by about 40% and sped up package imports by 21%.
Quantum computing gets leaner
Quantum computers require carefully designed operations to perform calculations while limiting errors and hardware demands. Some algorithms depend on smaller mathematical components, known as subroutines, that can consume substantial computing resources.
Gemini 4 Argon is here.
— Sundar Pichai (@sundarpichai) October 13, 2026
It’s available globally today for developers and enterprises in the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers in Google Antigravity, and Ultra and Pro subscribers in the Gemini app and Search.
Hope you all enjoy it! pic.twitter.com/2jyAgqoylU
Google said Argon optimized an arcsin subroutine used in quantum simulation algorithms. Compared with a published quantum CORDIC implementation, the new version reduced spacetime cost by more than 100 times.
Spacetime cost measures the computing resources needed to carry out an operation, accounting for both hardware requirements and execution time. The improvement applies to the specific subroutine tested, rather than quantum computing systems as a whole.
Argon also optimized an internal version of AlphaQubit, Google’s machine-learning-based quantum error correction decoder. By converting it into a single hardware-optimized kernel, the model nearly doubled inference throughput, according to Google.
The model’s software engineering capabilities were tested on Google’s open-source Agent Development Kit, a framework used to build and deploy autonomous agents.
During an 11-hour overnight run, Argon examined a codebase containing approximately 205,000 lines of code across nearly 800 files. It proposed six optimizations, two of which were accepted after human review and merged into the project.
The changes reduced in-memory session preparation overhead by about 40% during long sessions and improved package import speed by 21%.
In an independent benchmark, a 158-turn agent run fell from 65.6 seconds to 56.1 seconds. However, the time from starting a new process to its first model request remained essentially unchanged.
Fusion equations meet robotics
In nuclear fusion research, Google worked with University of Maryland professor Matt Landreman to identify a broader class of counterexamples to Grad’s conjecture, a mathematical hypothesis dating to 1967.
Earlier counterexamples had already challenged the conjecture. The newly identified solutions could provide reference cases for testing simulation software used to design stellarators, which use complex magnetic fields to confine hot plasma.
Separately, Argon demonstrated its ability to generate training environments for physical robots.
Starting with a single photograph of a tabletop, the model built a physics-based simulation and generated more than 4,500 training episodes for robotic manipulation tasks.
Gemini 4 Argon taught a robot to play cornhole just by watching a video.
— Google DeepMind (@GoogleDeepMind) October 13, 2026
It mapped the physics, coded its own simulation, and taught @Apptronik's Apollo humanoid to sink 7 throws in a row. pic.twitter.com/JIyoKnVc4q
Google reported that the resulting training data increased real-world robot success rates from 56% to 75%, while reducing task completion time from 35 to 24 seconds.
In another demonstration, Argon helped a Boston Dynamics Spot robot practice opening a foot-pedal trash bin in simulation before attempting the task with physical objects.
Beyond robotics, researchers at Gamow Labs used Argon to examine de-identified genetic data from newborns with rare developmental lung disorders. Google said the model identified a chromosomal rearrangement that conventional software had missed for years.
The results span scientific research, software engineering and physical robot control. However, the reported improvements remain specific to the experiments described, and broader performance across these fields has not yet been established.