AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Gemini 4 Argon: Built to solve your hardest problems

Collected Oct 13, 2026

Google announced Gemini 4 Argon, its newest frontier model, making it available globally on launch day across several surfaces: the Gemini API in AI Studio and the Gemini platform for developers and enterprises, Google Antigravity for Google AI Ultra subscribers, and the Gemini app plus AI Mode in Search for Ultra and Pro subscribers. Koray Kavukcuoglu, SVP of Google DeepMind and chief AI architect at Google, is named on the announcement. The launch carries an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95 percent.

The headline claim is quantitative: Google says Argon outperforms other frontier models on benchmarks across domains and ranks first on the Vals Index, which measures model performance across economic sectors including finance, coding, legal, and tax work. Google also reports an 87 percent score on the Embodied Reasoning (ERQA) benchmark, which tests understanding of physical and spatial problems in 3D environments, saying Argon leads all other frontier models there on pure reasoning.

Google describes Argon as strongest at complex, multi-step work in coding, finance, legal, cybersecurity, and science. For developers, the most concrete specification is scale: an industry-leading 1M unified input-output token limit via the Gemini API is what Google says enables the model to work longer on hard scientific, coding, and agentic problems. That unified input-output framing suggests the context budget is shared across what goes in and what comes out, which is worth checking against the API documentation before designing long-session agents around it. Google says the model communicates clearly, keeps users in the loop, follows style instructions better than prior Gemini models, answers more directly with less hedging, avoids repetition, and skips clichés.

Google's internal engineering results are the most concrete evidence offered. In Antigravity, agentic code submissions rose by more than 35 percent while code rollbacks fell by nearly a third. The Chrome team used Argon to separate core webpage handling from roughly 6 million lines of legacy browser feature code inside Chromium's 40 million-line codebase; Google says the refactor spanned hundreds of files and compressed months of work into weeks, and that the result will let live webpages move between tabs, side panels, and floating windows without reloading, eliminating a category of browser crashes. Separately, on Google's open-source Agent Development Kit (adk-python), Argon was given a skills rulebook and a one-paragraph goal, worked autonomously overnight, and surfaced optimizations now merged upstream that cut in-memory session prep time by 40 percent in long sessions and sped cold imports by 21 percent.

Open-source performance work recurs. Researchers used Argon in Antigravity to make the quantum chemistry package xTB up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search, with changes contributed to the official GitHub repository. On the open Gemma models, Google reports a 2.85x end-to-end inference speedup on an A100 via vLLM, taking Gemma 4 E2B from roughly 37,900 to over 108,000 tokens per second, with full token parity and no quality drop, achieved in days rather than months, with the changes upstreamed as pull requests spanning attention tuning and speculative decoding.

Argon is also positioned as a generator of visual and interactive software. Google reports improved spatial reasoning, mathematical precision, and visual coding, and describes Googlers prototyping 3D visualizations, high-fidelity games, and interactive animations in vanilla code that renders in the browser. Examples include a detailed Blender model of a pagoda produced through multi-agent visual critique loops; a playable miniature racing game built from photos of a real desk, with Blender MCP handling modeling and collision checks, Nano Banana texturing, and Three.js wiring for physics, shaders, and engine audio from Gemini Omni Flash; a Python-scripted 3D game rigged and animated with fragment-shader and physics work; and a pure-Python 3D ray tracer, written with no external libraries, that produced and narrated a video using Gemini 3.8 TTS and Lyria 3.5 music. Several of these are available to try in Google AI Studio.

Professional and enterprise case studies span document-heavy work. Google cites early-access customers doing deep multimodal document comprehension with strict grounding, longitudinal data analysis, and interactive analytical sandboxes. Examples include reconstructing 3D models of Chaco Canyon from roughly 50 GB of airborne LiDAR scans, GIS room vectors, Library of Congress CAD drawings, and National Park Service archives; turning an unstructured project folder into a formatted client proposal deck and then emailing it while setting up a scheduled workflow for future requests; and generating and 3D-printing replacement sofa legs sized from a sketch and a photo. Isomorphic Labs is using Argon on drug discovery research workflows, where Google says synthesizing complex PDFs and patents cuts a weeks-long process into hours.

On security, Google says Argon helps defenders audit critical codebases and formal proofs and catch zero-day vulnerabilities and logic flaws, and that it is expanding to hundreds of additional Fairwind Program partners starting launch day. Wiz used the model in its Scan for Good initiative: in one engagement with a federal agency it identified and remediated a gap risking exposure of sensitive data and government infrastructure, and in another it chained complex vulnerabilities to find and fix a public file download bypass that exposed a master key and granted administrator access.

Scientific results are unusually specific. In quantum error correction, Google says Argon translated an internal version of the AlphaQubit decoder into a single optimized hardware kernel, nearly doubling inference throughput to support larger models and devices under real-time thresholds. For an arcsin implementation used in quantum simulation, an Argon-written circuit verified in Lean cut spacetime cost by more than 100x against a published quantum CORDIC baseline. In fusion, Google and Professor Matt Landreman of the University of Maryland found a broader class of counterexamples to Grad's conjecture, the 1967 hypothesis that exact solutions for stellarator magnetic fields were impossible, giving the community new benchmarks. In rare disease research, Gamow Labs used Argon on de-identified newborn genomes with rare developmental lung disorders; Google says it pinpointed a chromosomal rearrangement that routine software missed for years, and that Argon topped Gamow Labs' RareBench (57 percent on rare-disease diagnosis assistance) and LabBench (52.2 percent on wet-lab reasoning). In enzyme design, Argon assembled data, tuned models, and applied biochemistry-informed filters to produce five novel serine esterase designs targeting 4MU-acetate, all laboratory-confirmed for catalytic activity within hours.

Google is offering 10,000 scientists the highest usage tier in Google Antigravity free for six months, onboarded on a rolling basis and subject to eligibility, capacity, and selection, with applications open now. On robotics, Google reports that Argon watched a raw video of a person playing cornhole, mapped the environment's spatial coordinates, computed whole-body joint movements, and autonomously wrote the simulation code; transferred to Apptronik's Apollo humanoid, it sank seven consecutive shots. Google frames this video-to-sim-to-real pipeline as compressing weeks of manual modeling into hours.

Safety work is presented as the most comprehensive Gemini stack to date. Under the Frontier Safety Framework, Google deployed layered jailbreak resistance, monitoring, and enforcement; external bio red teaming unlocked less than 1 percent of tested bio sub-tasks, and violations on an adversarial stress set fell nearly two-fold versus previous Gemini models. Argon was trained to respect system boundaries, report negative findings honestly, and pause for human confirmation before irreversible actions, with shortcut attempts on coding benchmarks dropping to 2 percent. An automated misalignment monitor reads agent reasoning and actions to block workarounds in real time, without penalizing the model's reasoning during training. Against indirect prompt injection, Google says Argon leads the Gray Swan benchmark, cutting attack success across 15 adaptive attempts to 0.7 percent. A model card and developer documentation accompany the release.

Why it matters: teams evaluating frontier models now have a directly comparable price point, and the launch-day availability across API, Antigravity, and consumer surfaces means developers can test Argon against incumbent models immediately rather than waiting. The reported gains on refactoring, framework overhead, and inference kernels suggest the strongest near-term use cases are agentic code work on large codebases and performance engineering, though those figures come from Google's own engineering teams and should be reproduced independently. Product teams should also weigh the safety posture, particularly the 0.7 percent prompt injection success rate and the 2 percent shortcut attempt rate, as inference rather than guarantee, since these numbers reflect Google's own adversarial sets.

Read at Google AI Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

TBD We’re giving developers, enterprises, scientists, and consumers access to Gemini 4 Argon, our latest frontier model.