Google DeepMind now lists Gemini 4 Argon as available to Google AI Pro and Ultra subscribers. The model can be used through Gemini, Google AI Studio and Google Antigravity. The page advertises a 1M-token output window for long derivations, repository-wide refactoring and large structured outputs.

Google reports several software engineering results. Argon helped the Chrome team separate a roughly 6 million line legacy folder from Chromium’s 40 million line codebase. The refactor covered hundreds of files and reduced work estimated at months to several weeks. Google says the change will support live movement of webpages between tabs, side panels and floating windows without reloads, while removing a category of browser crashes.

In the xTB quantum chemistry package, Argon reportedly improved molecular-structure prediction by up to 2.3x and conformer searches by 1.8x. The changes were submitted to the official xTB repository. Work on Google’s open-source Agent Development Kit, guided by a .agents/skills rulebook, cut long-session in-memory preparation time by 40% and cold imports by 21%.

A separate vLLM effort on an A100 raised Gemma 4 E2B inference throughput from about 37,900 to more than 108,000 tokens per second. Google reports a 2.85x end-to-end gain, 100% token parity and no quality loss. Argon also migrated more than 700 Google Skills training labs across five repositories, found at least 16 timeout and race-condition bugs, saved 955 engineering hours and reduced recurring maintenance from about 540 hours per year to one three-hour update.

The demonstrations include multi-agent Blender work, a desk converted into a miniature racing game, and a 3D game assembled with Python, Three.js and GLSL shaders. In one example, Argon generated a pure-Python ray-tracing engine without external libraries. The resulting video used Gemini 3.8 TTS and music from Lyria 3.5. Another workflow reconstructed Chaco Canyon from about 50 GB of LiDAR, GIS, CAD and archive data.

Google and its partners report gains in professional workflows. Isomorphic Labs uses Argon to synthesize patents and complex PDFs for drug discovery, reducing work that usually takes weeks to hours. Ambience Healthcare found high recall and ICD-10 specificity on more than 200 synthetic longitudinal charts and is testing the system on real patient records. Box reports a 14% accuracy improvement over Gemini 3.1 Pro. Cartwheel says its spatial reconstruction pipeline runs nearly 100 times faster and uses 90% fewer tokens.

Deloitte reports a 100% completion rate in early evaluations of long-running, tool-using workflows. Kepler says Argon matched analyst-built financial models on more than 99% of figures. Quadrillion reports frontier performance on complex research tasks at 1.5 times the speed of peer models and one quarter of their per-task cost. LexisNexis cites early strength in legal research and drafting. MinutalA says the model found a legally relevant detail missed by a human answer key. Roboflow has integrated Argon after it ranked first in its object-detection evaluations and led a text benchmark based on industrial and aerial imagery.

The research examples cover quantum computing, fusion, rare-disease diagnosis and enzyme design. An Argon-optimized AlphaQubit kernel nearly doubled inference throughput. A Lean-verified implementation of arcsin reduced spacetime cost by more than 100x against a published quantum CORDIC baseline. Google and Matt Landreman found a broader class of counterexamples to Grad's conjecture, creating benchmarks for stellarator simulations. Gamow Labs reports that Argon identified a chromosomal rearrangement missed by routine software, scoring 57% on RareBench and 52.2% on LabBench. Five new serine esterase designs showed catalytic activity in laboratory tests.

In robotics, Argon watched a video of a human playing cornhole, built a physical simulation and calculated whole-body movements. An Apptronik Apollo humanoid then made seven consecutive successful shots. In cybersecurity, the model scored 68% on CWE-bench v1, tied for first in the published comparison, and Google calls it its most resilient model against indirect prompt injection. Wiz used it in the Scan for Good program to protect healthcare infrastructure and hundreds of thousands of patient records.

The published evaluations report 68.9% on Vals Index, 77.9% on DeepSWE v1.1, 55.0% on FrontierSWE v2 and 91.9% on Vibe Code Bench. Argon scored 88.8% on LABBench and 76.0% on RiemannBench. Its long-context score was 99.7% at 128K tokens and 84.2% from 256K to 1M. The table also reports 39.5% on Agent's Last Exam, 69.2% on the offline OSWorld-2.0 subset, 71.6% on Chartography and 91.7% on LVBench.

Google says Argon refuses harmful requests under its Frontier Safety Framework. Monitoring can stop execution when misalignment signals appear, and sandbox environments are isolated before high-risk training or evaluation. Academic and independent researchers can apply for six months of the highest Google Antigravity usage tier, subject to eligibility, capacity and selection. Cybersecurity defenders can apply to the Fairwind Program.