Google Launches Gemini 4 Argon for Complex Reasoning and Code Refactoring

Gemini 4 Argon is Google’s newly released high-precision, multi-domain reasoning model, now available globally via API with launch pricing of $2 per million input tokens and $10 per million output tokens.

Argon leads in complex tasks across finance, programming, law, cybersecurity, and science, ranking first on the Vals Index. Internally at Google, it has increased agentic code submissions by over 35% and reduced rollback rates by nearly one-third. Cached input tokens are eligible for a 95% discount.

This release is ideal for AI platform engineers, enterprise LLM application architects, and vertical domain (finance/law/security/research) AI engineers.

Today, Gemini 4 Argon is open to developers and enterprises worldwide via the Gemini API, and is also available to Google AI subscribers¹ within Google Antigravity, the Gemini App, and Search’s AI Mode.

Argon excels at deep work on complex real-world tasks in programming, finance, law, cybersecurity, and science. Its depth and precision in multi-step tasks allow it to outperform other frontier models across various benchmarks, securing the top rank on the Vals Index, which measures model performance in key economic sectors such as finance, programming, law, and tax.

Launched today, the introductory price² is $2 per million input tokens and $10 per million output tokens (with a 95% discount for cached input tokens).

Google internal teams, enterprises, and the research community have already begun using Argon, achieving significant breakthroughs across multiple domains:

• Solving major engineering challenges like large-scale codebase refactoring and optimization
• Optimizing cross-domain business workflows in law, finance, and science
• Discovering and fixing critical cybersecurity vulnerabilities
• Driving innovative breakthroughs in quantum computing, nuclear fusion, and life sciences

Solving Complex Engineering Tasks

At Google, we have been using Argon to tackle our most difficult engineering problems. Over the past few weeks, software engineers have delivered high-quality code faster than ever before. Argon within Antigravity helped teams increase agentic code submissions by more than 35%, while reducing code rollback rates by nearly one-third.

• Untangling browser feature code in Chrome: Argon helped the Chrome team separate core web processing code from a legacy browser feature folder containing approximately 6 million lines of code, buried within Chromium’s 40-million-line codebase. By batching multiple clean code change optimizations, Argon assisted the team in completing an architectural refactor involving hundreds of files, shrinking months of work into just a few weeks. For billions of Chrome users worldwide, this enables loaded pages to move seamlessly between tabs, sidebars, and floating windows without reloading, resolving an entire class of issues that cause browser crashes.

• Optimizing production-grade Python codebases in hours: Argon in Antigravity is helping the team behind Google’s Agent Development Kit (ADK)—a framework powering enterprise-grade agents—further improve its performance. The team provided Argon with their .agents/skills rulebook and a clear, concise goal: eliminate framework overhead without breaking existing applications. Working autonomously overnight, Argon proposed various optimization schemes (now live in google/adk-python), reducing memory session preparation time in long sessions by 40% and improving cold import speeds by 21%.

• Optimizing the open-source quantum chemistry package xTB: xTB (Semiempirical Extended Tight-Binding) is one of the most cited open-source quantum chemistry software packages in the molecular field. Scientists use it to predict 3D molecular structures and screen vast numbers of candidate drugs and materials. Using Argon in Antigravity, our researchers improved xTB’s molecular structure prediction speed by 2.3x and conformational search speed by 1.8x. These optimization updates have been contributed to the official xTB repository on GitHub.

• Boosting our Gemma open model speed by nearly three times: With Argon, we achieved a 2.85x end-to-end inference speedup on A100 GPUs via vLLM in just a few days (accelerating throughput for Gemma 4 E2B from ~37,900 tokens/sec to over 108,000 tokens/sec), ensuring 100% output token consistency and zero quality degradation—a process that typically takes months. Check out the upstream pull requests (covering comprehensive optimizations from attention mechanism tuning to speculative decoding) for specific technical details.

• Implementing complex 3D visualizations: Compared to previous Gemini models, Argon delivers significant improvements in spatial reasoning, mathematical precision, and visualization programming. By using commands like /goal and initiating multi-agent loops in Antigravity with custom skill prompts, Google employees are leveraging Argon to create complex 3D visualizations, high-fidelity games, and interactive animation prototypes directly via native code, all renderable in the browser.

Transforming Key Business and Professional Workflows

Argon combines frontier reasoning, deep visual perception, expanding domain knowledge, and stronger writing capabilities in law, finance, science, and programming to assist with high-stakes professional workflows. It communicates clearly, keeping you updated on progress, and adheres to style instructions more reliably than any previous Gemini model. Argon answers questions more directly, avoids ambiguity, eliminates unnecessary repetition, and steers clear of clichés in its responses.

Enterprise users and startups in our early access program have already achieved notable results with Argon. Leveraging its advanced reasoning and multimodal capabilities while maintaining strict factual grounding, Argon excels in deep multimodal document understanding, longitudinal data analysis, and interactive analysis sandboxes.

Accelerating Real-World Change and Impact with Researchers and Industry Experts

Researchers, engineers, and security experts are already using Argon to push the boundaries of cyber defense, science, mathematics, and robotics. It is helping cyber defenders and researchers validate hypotheses, autonomously audit critical codebases and formal proofs, and uncover elusive zero-day vulnerabilities and logic flaws.

Bringing Advanced Cybersecurity Capabilities to More Trusted Defenders

Starting today, we are opening the "Fairwind Program" to hundreds of new partners.

Wiz has already deployed Argon in its "Scan for Good" initiative, aimed at protecting critical infrastructure in the public interest sector for free. By gaining early access to the model, Wiz demonstrated Argon’s real-world impact across multiple high-risk domains, including critical healthcare infrastructure. Through the analysis of complex vulnerability chains, the model helped teams discover and rapidly patch a public file download bypass vulnerability in medical software (which had previously led to master key leaks and privilege escalation for attackers). Argon did not just find vulnerabilities; it actively protected hundreds of thousands of private patient records.

Unlocking New Scientific Breakthroughs

Argon is already making a tangible impact in the scientific community, enabling scientists and experts across various fields to achieve innovative breakthroughs.

• Quantum Computing: Argon is helping our researchers accelerate breakthroughs in quantum error correction and algorithm design. To ensure error-correction decoders keep pace with quantum hardware in real time, Argon converted an internal version of our machine learning-based quantum error correction decoder, AlphaQubit, into highly optimized hardware kernels. This nearly doubled inference throughput, allowing us to deploy larger, more capable models and scale to larger quantum devices while continuing to meet strict real-time threshold requirements. Argon also reduced the space-time cost of quantum subroutines. Taking the arcsin primitive from recent quantum simulation algorithms as an example, the implementation written by Argon and formally verified in Lean achieved over 100x lower space-time costs compared to published quantum CORDIC benchmarks. This marks a significant step toward optimizing standard libraries for quantum circuits.
• Nuclear Fusion: Stellarators are extremely complex machines and leading candidates for achieving infinite clean fusion energy, but optimizing their 3D magnetic fields has historically been challenging. Since 1967, a long-standing hypothesis known as the "Grad Conjecture" posited that finding exact mathematical solutions for these magnetic fields was impossible. Recently, researchers, including Professor Matt Landreman from the University of Maryland, discovered counterexamples. Now, the Google team, collaborating with Professor Landreman, has found additional counterexamples to the Grad Conjecture. These new solutions provide the nuclear fusion community with reliable benchmarks for testing their stellarator simulation codes—tools that are invaluable for future designs.
• Rare Disease Research: Researchers at Gamow Labs used Argon to analyze de-identified genomes from infants suffering from rare developmental lung diseases. In one case, Argon precisely identified a chromosomal rearrangement that conventional software had missed for years and explained why it went undetected. Internal evaluations by Gamow Labs, based on real clinical and laboratory data, showed that Argon ranked first on both RareBench (achieving 57% in assisting rare disease diagnosis) and LabBench (achieving 52.2% in wet-lab reasoning). These advances suggest that families facing rare diseases will receive diagnoses faster.
• Enzyme Design: Designing enzymes with precise reaction activity has traditionally been time-consuming, often requiring months of custom model fine-tuning and candidate screening. Serine esterases are a class of enzymes that cleave ester bonds on target molecules and are widely used in recycling, as well as in the manufacturing of pharmaceuticals, biofuels, and detergents. Our scientists collaborated with Argon to design serine esterases targeting the benchmark molecule 4MU-acetate, which fluoresces under UV light after its ester bond is cleaved. Within hours, Argon gathered relevant data, completed model fine-tuning, and identified multiple screening rules based on biochemical principles, ultimately yielding five novel design combinations. Subsequent laboratory tests confirmed that all these designs exhibited strong catalytic activity. This successful case of rapid and accessible enzyme generative design demonstrates Argon’s immense potential to accelerate enzyme engineering processes in fields such as biomanufacturing, green chemistry, and drug development.

To further accelerate research progress, we will provide 10,000 scientists with six months of free access to the highest usage tier of Google Antigravity (via rolling approval), subject to eligibility, capacity, and screening. Applications are now open.

Accelerating Robotics Development

By combining advanced visual semantic reasoning, deep physical understanding, and autonomous code generation, Argon unlocks new workflows across the entire robotics stack and accelerates the path toward Physical AGI. This leap in capability is reflected in Argon’s outstanding performance on the Embodied Reasoning QA (ERQA) benchmark, which evaluates a model’s ability to understand and solve physical and spatial problems in 3D environments. Relying solely on reasoning capabilities, Argon scored 87%, surpassing all other frontier models.

To demonstrate these capabilities in practice, we tested Argon’s ability to learn out-of-domain physical skills simply by watching. In an internal test, Argon watched raw video footage of humans playing bean bag toss, mapped the spatial coordinates of the environment, and calculated the required full-body joint movements. It then autonomously constructed simulation code. When this code was transferred to Apptronik’s Apollo humanoid robot, the robot successfully hit seven consecutive targets in real-world trials.

This "Video-to-Simulation-to-Reality" pipeline compresses weeks of manual simulation modeling into hours, transforming ordinary videos into scalable, infinite training grounds for robots. To learn how this technology accelerates robotics development, visit our developer blog.

Developing Argon Responsibly and Safely

Safety and security are foundational to Google’s model development. Given Argon’s exceptional performance, we have collaborated with numerous third-party partners to evaluate specific model capabilities and safety safeguards. Gemini 4 Argon represents our most comprehensive safety and security framework to date:

• Defense Against Abuse and Jailbreaks: We empower defenders with robust programming and scientific capabilities to patch medical software vulnerabilities or accelerate bioengineering modifications, while simultaneously implementing strong safeguards against malicious actors. Guided by our Frontier Safety Framework, we evaluated Argon’s enhanced capabilities across multiple domains and deployed a multi-layered defense system combining jailbreak resistance, monitoring, and enforcement, significantly improving model robustness. For instance, in external biosafety red-team tests, fewer than 1% of bio-related subtasks were successfully induced for execution. Furthermore, regarding dangerous requests (ranging from writing exploit code to assisting in the acquisition of hazardous biological materials), Argon demonstrated significantly better performance in refusing harmful requests compared to our previous models. In an adversarial test set designed to stress-test the model, the violation rate decreased by nearly 2x—again, significantly outperforming prior Gemini models.
• Agentic Alignment and Real-Time Monitoring: As agents’ problem-solving capabilities improve, we must ensure they do not take unauthorized shortcuts when encountering obstacles such as permission errors or test failures, driven by a rush to complete tasks.

First, we trained Argon to respect system boundaries, honestly report negative results, and pause for human confirmation before taking irreversible steps. In coding benchmarks where the model could operate without constraints, the shortcut attempt rate for Gemini 4 Argon dropped to just 2%, representing our best performance relative to previous Gemini models.

Second, as a defense-in-depth layer, we built an automated deviation monitor that reads the agent’s reasoning process and specific actions to capture and intercept unintended workarounds in real time. We used this monitor during internal deployments, capability evaluations, and reinforcement learning training runs, while strictly avoiding penalizing the model’s thought process during training to keep its reasoning transparent and monitorable.

• Prompt Injection Resistance: When AI agents browse the web, read emails, or inspect code, they may encounter hidden malicious instructions designed to hijack their tasks. Through adversarial training and automated red-teaming, Argon leads on the Gray Swan Indirect Prompt Injection (IPI) benchmark, reducing the attack success rate over 15 adaptive attempts to 0.7%.