Google DeepMind launches Gemini 4 Argon for complex coding, science and enterprise tasks

Google DeepMind has announced the global availability of Gemini 4 Argon, a frontier model designed for complex, multi-step work in coding, finance, legal services, cybersecurity and science.

The model is available from Oct. 13, 2026, to developers and enterprises through the API in AI Studio and the Gemini platform. Google AI Ultra subscribers can use it in Google Antigravity, while Google AI Ultra and Pro subscribers can access it in the Gemini app and AI Mode in Search.

Koray Kavukcuoglu, SVP of Google DeepMind and chief AI architect at Google, said the release gives developers, enterprises, scientists and consumers access to the company’s latest frontier model.

Google said Argon is designed for “deep work” across real-world domains. The company said the model outperforms other frontier models on benchmarks across multiple fields and ranks No. 1 on the Vals Index, which measures model performance in critical economic sectors including finance, coding, legal work and tax work.

The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. After the introductory period ends on Nov. 30, 2026, the price will be $4 per 1 million input tokens and $20 per 1 million output tokens. The pricing applies to the Argon model intended for public access.

Google said teams are already using Argon to refactor and optimize large codebases, improve workflows in legal, finance and science, identify and patch cybersecurity vulnerabilities, and support research in quantum computing, nuclear fusion and life sciences.

Engineering and software development

Google said Argon has been used internally on difficult engineering projects. In Antigravity, the model helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly one-third.

At Chrome, Argon helped engineers separate core webpage handling from a legacy folder containing about 6 million lines of browser features within Chromium’s 40 million-line codebase. By producing multiple clean code changes at once, the model helped complete an architectural refactor spanning hundreds of files in weeks rather than months.

Google said the work is intended to allow live webpages to move between tabs, side panels and floating windows without reloading, eliminating a category of browser crashes for Chrome’s billions of users.

Argon was also used by the team behind Google’s open-source Agent Development Kit, or ADK, a framework for enterprise agents. Engineers provided the model with their .agents/skills rulebook and a one-paragraph objective: eliminate framework overhead without breaking existing applications.

Working autonomously overnight, Argon identified optimizations that are now live in the google/adk-python repository. Google said the changes reduced in-memory session preparation time by 40% during long sessions and improved cold-import speed by 21%. The company published an engineering deep dive describing the work.

Researchers also used Argon in Antigravity to optimize xTB, or Semiempirical Extended Tight-Binding, an open-source quantum chemistry package used to predict molecular 3D structures and screen drug and materials candidates. Google said xTB became up to 2.3 times faster at predicting molecular structures and 1.8 times faster at conformer searches. The changes were contributed to the official xTB GitHub repository.

Google also reported a 2.85-times end-to-end inference speedup for its open Gemma models on an A100 using vLLM. Throughput for Gemma 4 E2B increased from about 37,900 to more than 108,000 tokens per second, with 100% token parity and no reduction in quality. Google said the work, which included attention tuning and speculative decoding, was completed in a few days rather than the months normally required. The associated upstream vLLM pull requests contain the changes.

3D visualization and interactive applications

Google said Argon improves spatial reasoning, mathematical precision and visual coding compared with earlier Gemini models. Using slash commands such as /goal, multi-agent loops and custom skills in Antigravity, Google employees have used the model to prototype 3D visualizations, high-fidelity games and browser-rendered interactive animations using vanilla code.

In one demonstration, Argon created a detailed pagoda in Blender through multi-agent visual-critique loops. The model autonomously sculpted a multi-building temple complex with photorealistic architecture, foliage and lighting.

In another example, Argon converted photographs of a real desk into a playable miniature racing game. Blender MCP was used to model and collision-check the desk, room and racetrack; Nano Banana supplied textures; and Three.js handled the physics, custom shaders and car-engine audio from Gemini Omni Flash. Google made the application available through Google AI Studio.

Argon also wrote Python scripts to procedurally sculpt, rig, animate and physically based render a complex 3D game. It then connected the project to the web in Three.js with GLSL shaders, shatter physics and a single-player AI. The resulting application is available in Google AI Studio.

Using no external libraries and 100% pure Python, Argon created a 3D ray-tracing engine with a custom video tool and audio mixer. The system built and documented a sunlit library from scratch, producing, narrating and assembling a video with Gemini 3.8 TTS voiceover and music by Lyria 3.5. Google published the source code.

In another Antigravity demonstration, Argon began with a single Nano Banana image and used Blender MCP in an autonomous loop to align camera grids, sculpt Gothic architecture and stage audio-reactive lighting for a finished 3D scene.

Business and professional workflows

Google said Argon combines advanced reasoning, visual perception, broader domain knowledge and stronger writing capabilities for high-stakes professional workflows. The company said the model communicates clearly, follows style instructions more consistently than previous Gemini models, answers more directly with less hedging, avoids unnecessary repetition and omits clichés.

Enterprise customers and startups in Google’s early-access program are using Argon for multimodal document comprehension, longitudinal data analysis and interactive analytical sandboxes while maintaining strict grounding, according to Google.

One demonstration involved roughly 50 GB of Airborne LiDAR scans, GIS room vectors, Library of Congress CAD drawings and National Park Service archives. Argon synthesized the material to reconstruct 3D models of Chaco Canyon and turn fragmented archaeological records into explorable digital heritage sites. The project is available through Google AI Studio.

In the Gemini app, Argon can take an unstructured folder containing project context and format it into a proposal deck for a client using a custom skill. Gemini can then draft and email the deck while creating a scheduled workflow for future incoming client requests.

Google also demonstrated Argon combining data from three open-source repositories to produce a medical learning tool. The tool allows users to peel back anatomical layers, inspect clinical metadata and trigger biomechanics in a 3D exploration of the human body.

In the Gemini macOS app, Argon used Antigravity and Nano Banana to generate design iterations for new sofa legs and then create a 3D-printable file. Starting with a measurement sketch and a photograph of the couch, the model built the part to the correct size.

At Isomorphic Labs, Argon is being used to accelerate drug-discovery and design research. Google said the model can synthesize complex PDFs and patents to extract relevant information, reducing a process that typically takes weeks to several hours.

Cybersecurity and scientific research

Researchers, engineers and security specialists are using Argon in cyber defense, science, mathematics and robotics, Google said. Applications include testing hypotheses, auditing critical codebases and formal proofs, and identifying zero-day vulnerabilities and logic flaws.

Fairwind Program and Wiz

Google said it is beginning an expansion to hundreds of additional partners in the Fairwind Program, which provides cyber capabilities to trusted defenders.

Wiz is using Argon in its “Scan for Good” initiative, which protects critical infrastructure for the public good at no cost. According to Google, Wiz used early access to Argon in multiple high-stakes areas.

In a long-running cyber-resilience program conducted with a federal agency, Argon helped identify and remediate a gap that risked exposing sensitive government data and infrastructure. Google said the effort helped prevent exposure of critical security-sector development details and unauthorized access to specialized models.

Wiz also used Argon to protect healthcare infrastructure. By chaining complex vulnerabilities, the model helped identify and quickly fix a public file-download bypass that exposed a master key and granted administrator access. Google said Argon went beyond identifying the vulnerability and helped protect hundreds of thousands of private patient records.

Quantum computing

Google researchers used Argon to support work in quantum error correction and algorithm design. To help error decoders keep pace with quantum hardware in real time, Argon translated an internal version of AlphaQubit, Google’s machine-learning quantum error-correction decoder, into a single optimized hardware kernel.

Google said the kernel nearly doubled inference throughput, allowing the deployment of larger models and scaling to larger quantum devices while maintaining strict real-time thresholds.

Argon was also used to reduce spacetime costs for quantum subroutines. For arcsin, a building block in recent quantum simulation algorithms, an implementation written by Argon and formally verified in Lean reduced spacetime cost by more than 100 times compared with a published quantum CORDIC baseline. Google described the work as a step toward a standard library of optimized quantum circuits. Related research is available in an arXiv paper.

Nuclear fusion

Stellarators are complex machines and a leading candidate for clean fusion energy, but optimizing their three-dimensional magnetic fields is difficult. Since 1967, a hypothesis known as “Grad’s conjecture” held that finding exact mathematical solutions for these fields was impossible. Recent counterexamples, including work by Professor Matt Landreman of the University of Maryland, challenged that assumption.

Google said a team working with Landreman has now identified a broader class of counterexamples to Grad’s conjecture. The new solutions provide benchmarks for the fusion community to use when testing stellarator simulation codes, which support future designs.

Rare-disease research

Researchers at Gamow Labs used Argon to analyze de-identified genomes from newborns with rare developmental lung disorders. In one case, the model identified a chromosomal rearrangement that routine software had missed for years and explained why it had gone undetected.

On private Gamow Labs evaluations derived from real clinical and laboratory data, Argon scored 57% on RareBench for assistance with rare-disease diagnosis and 52.2% on LabBench for wet-lab reasoning. Google said the results could support faster diagnosis for families facing rare conditions. Additional information is available from Gamow Labs and its benchmarks page.

Enzyme design

Engineering enzymes with precise reactivity often requires months of model tuning and candidate screening. Serine esterases, which cut ester bonds in targeted molecules, are used in recycling and in the manufacture of pharmaceuticals, biofuels and detergents.

Google scientists worked with Argon to design serine esterases targeting 4MU-acetate, a benchmark molecule that glows under ultraviolet light when its ester bonds are broken. In a matter of hours, Argon assembled relevant data, performed model tuning and identified several biochemistry-based filters, producing a final set of five novel designs.

Subsequent laboratory testing confirmed robust catalytic activity across the designs. Google said the work demonstrates the potential for faster enzyme engineering in biomanufacturing, green chemistry and drug development.

Google is offering 10,000 scientists access to the highest usage tier in Google Antigravity free of charge for six months. Researchers will be onboarded on a rolling basis, subject to eligibility, capacity and selection. Applications are open.

Robotics

Google said Argon combines visual-semantic reasoning, physical understanding and autonomous code generation to support new workflows across the robotics stack. On the Embodied Reasoning, or ERQA, benchmark, which tests a model’s ability to understand and solve physical and spatial problems in 3D environments, Argon achieved an 87% score. Google said this was higher than the scores of other frontier models when relying on pure reasoning. The ERQA benchmark is available publicly.

In an internal test, Argon learned an out-of-distribution physical skill by watching a raw video of a person playing cornhole. The model mapped the environment’s spatial coordinates, calculated the required whole-body joint movements and autonomously built simulation code.

When the system was transferred to Apptronik’s Apollo humanoid, the robot made seven consecutive shots in a real-world trial. Google called the process a “Video-to-Sim-to-Real” pipeline and said it compresses weeks of manual simulation modeling into several hours by turning ordinary videos into a scalable training environment for robotics. Apptronik is the company behind the Apollo humanoid. Google provided additional details in its developers blog.

Safety and security

Google said safety and security are foundational to its model-development process. Because Argon is highly capable, the company worked with third-party partners to measure certain capabilities and safeguards. Google described Gemini 4 Argon as having its most comprehensive safety and security stack to date.

Under Google’s Frontier Safety Framework, the company evaluated Argon’s capabilities in several areas and deployed a multilayer defense system combining jailbreak resistance, monitoring and enforcement. Google said external biological red teams were able to unlock less than 1% of the biological sub-tasks tested.

For dangerous requests, including exploit-code generation and assistance involving dangerous biological materials, Google said Argon was significantly less likely than previous models to comply. On an adversarial test set designed to stress the model, violations fell by nearly twofold compared with previous Gemini models.

Google also trained Argon to respect system boundaries, report negative findings honestly and pause for human confirmation before taking irreversible actions. On coding benchmarks where models had no incentive to hold back, shortcut attempts fell to 2%, which Google described as its best result compared with previous Gemini models.

As an additional defense layer, Google built an automated misalignment monitor that reads an agent’s step-by-step reasoning and actions to identify and block unintended workarounds in real time. The monitor was used in internal deployments, capability evaluations and reinforcement-learning training runs. Google said it avoided penalizing the model’s thoughts during training so that its reasoning would remain transparent and monitorable.

Argon was also tested against prompt injection. Agents that browse webpages, read emails or inspect code can encounter hidden malicious instructions intended to hijack their tasks. Through adversarial training and automated red teaming, Argon leads the Gray Swan Indirect Prompt Injection benchmark, with attack success reduced to 0.7% across 15 adaptive attempts.

Further information is available in the Gemini 4 Argon Model Card.

Availability and access

Argon supports an industry-leading 1 million-token unified input-output limit through the Gemini API. Google said developers and enterprises should review its documentation on the new continuation protocol and configure their systems appropriately. Documentation is available for long decoding and Gemini 4 Argon on Google’s enterprise agent platform.

Google’s global rollout includes the following access options:

• Developers: Build with the paid Gemini API tier in AI Studio and Android Studio. Google AI Ultra subscribers can also try Argon in Google Antigravity and AI Studio.
• Enterprises: Use Argon in Gemini for business, through the API and in Google Antigravity.
• Everyday users: Google AI Ultra and Pro subscribers can use Argon in the Gemini app and AI Mode in Search.
• Scientists and researchers: Apply for six months of free access to Google Antigravity’s highest usage tier, subject to eligibility, capacity and selection.
• Cybersecurity defenders: Apply to join the Fairwind Program for rolling access to Argon without cyber guardrails.

Google said Gemini 4 Argon is the first release in the Gemini 4 family and that additional models will follow.

The company noted that the views in the blog are those of the external party and do not represent Google’s views. The original announcement was published by Google DeepMind.