Google has announced the global availability of Gemini 4 Argon, a new AI model designed to handle complex, multi-step tasks across software engineering, finance, legal services, cybersecurity and scientific research.
The model is now available to developers and enterprise customers through the Gemini API in Google AI Studio and Google Cloud’s Gemini platform. Eligible Google AI Ultra subscribers can also access it through Google Antigravity, while Google AI Ultra and Pro subscribers can use it in the Gemini app and AI Mode in Search.
According to Google, Argon is built for demanding real-world tasks that require deeper reasoning, technical precision and the ability to work through multiple stages of a problem. The company says its performance across several professional domains places it ahead of competing frontier AI models on selected benchmarks.
One notable achievement is its reported first-place ranking on the Vals Index, an evaluation that measures AI performance across economically significant fields, including coding, finance, legal work and taxation.
Gemini 4 Argon Pricing and Availability
Google is introducing Gemini 4 Argon at a starting price of $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount, potentially reducing costs for applications that repeatedly process the same information.
The pricing structure makes the model accessible to developers building AI-powered applications, as well as businesses looking to integrate advanced reasoning capabilities into existing workflows.
Google Reports Major Improvements in Software Engineering
Google says its internal engineering teams have already used Argon to accelerate software development, improve code quality and tackle complicated projects that would traditionally require weeks or months of work.
When deployed through Google Antigravity, the model has reportedly helped teams increase agent-driven code submissions by more than 35%, while reducing code rollbacks by nearly one-third.
Several projects demonstrate how the company is applying the model to real engineering challenges.
Refactoring Chrome’s Massive Codebase
One of the most significant examples involves Chromium, the open-source browser project behind Google Chrome.
Google says Argon helped engineers separate core webpage-handling functionality from a legacy collection of browser features containing approximately six million lines of code. The wider Chromium codebase contains around 40 million lines.
By coordinating multiple code changes across hundreds of files, the model helped the engineering team complete an architectural refactoring project in weeks rather than months.
Google says the changes could eventually enable webpages to move between browser tabs, side panels and floating windows without requiring a reload. The improvements are also expected to address a category of browser crashes.
Google Uses AI to Optimize a Large Python Codebase
Google’s Agent Development Kit (ADK), an open-source framework used to build AI agents, has also benefited from Argon’s coding capabilities.
The engineering team gave the model its development rules and a high-level objective: reduce framework overhead without breaking applications that already depend on the software.
According to Google, Argon worked autonomously overnight to identify and implement performance optimizations.
The resulting changes reportedly reduced in-memory session preparation time by 40% during long sessions and improved cold-import speeds by 21%.
The optimizations have been contributed to the project’s public GitHub repository, demonstrating how AI-assisted engineering can produce changes that developers can inspect and use.
AI Accelerates Quantum Chemistry Research
Argon is also being applied to scientific computing.
Researchers used the model to optimize xTB, an open-source quantum chemistry package used to predict molecular structures and screen potential candidates for drug development and advanced materials.
Google reports that the optimizations made molecular structure predictions up to 2.3 times faster and conformer searches up to 1.8 times faster.
These improvements could help researchers process molecular candidates more efficiently, although the practical benefits will depend on their specific workloads and computing environments.
Making Gemma AI Models Nearly Three Times Faster
Google says Argon has also helped optimize inference performance for its Gemma open AI models.
In a test involving Gemma 4 E2B running through vLLM on an NVIDIA A100 GPU, the company reports a 2.85-fold improvement in end-to-end inference performance.
The reported throughput increased from approximately 37,900 to more than 108,000 tokens per second, with Google claiming identical token outputs and no reduction in quality.
The work involved several optimization techniques, including improvements to attention processing and speculative decoding. Google says the results were achieved within days, rather than the months such engineering efforts might otherwise require.
Stronger Reasoning for Professional and Business Workflows
Beyond software development, Gemini 4 Argon is designed to support demanding professional tasks across finance, law, science and other specialized fields.
Google highlights improvements in visual understanding, reasoning, domain knowledge and writing. The company also says the model follows instructions more consistently, communicates more directly and produces less repetitive output than previous Gemini models.
These capabilities could make Argon useful for businesses that need AI systems to analyze complex documents, investigate technical problems, streamline internal processes and assist professionals with research-intensive work.
However, applications involving legal decisions, financial judgments or other high-stakes matters will still require appropriate human oversight and verification.
Gemini 4 Argon Expands AI-Powered Cybersecurity
Cybersecurity is another major area where Google expects the model to have a significant impact.
The company says Argon can help security specialists audit complex codebases, investigate subtle software flaws, test hypotheses and identify vulnerabilities that may be difficult to detect through conventional approaches.
Google is also expanding access to the model through additional partners in its Fairwind Program, which supports trusted cybersecurity professionals working to strengthen digital defenses.
Helping Protect Government Systems
Cybersecurity company Wiz has been using Argon through its Scan for Good initiative, which focuses on identifying and addressing security exposures affecting critical infrastructure.
In one reported case involving a federal agency, Argon helped identify and remediate a security gap that could have exposed sensitive government data and infrastructure.
Google says the work helped prevent potential exposure of critical security-sector development details and unauthorized access to specialized models.
Securing Healthcare Infrastructure
Wiz has also used Argon to investigate vulnerabilities affecting healthcare infrastructure.
According to Google, the model helped security researchers connect multiple weaknesses to uncover a public file-download bypass. The issue exposed a master key and could have enabled administrator-level access.
The team subsequently addressed the vulnerability, helping protect sensitive systems and hundreds of thousands of private patient records.
The example highlights the potential value of advanced AI in identifying vulnerabilities that emerge when several weaknesses interact, rather than examining each issue in isolation.
What Gemini 4 Argon Means for the AI Industry
Google’s launch signals a continued shift in AI development from systems primarily focused on generating text and answering questions towards models designed to complete complex tasks and support professional workflows.
For software engineers, that could mean faster debugging, code optimization and large-scale refactoring. For businesses, it could mean more efficient research and process automation. For cybersecurity teams, advanced reasoning could help uncover weaknesses before malicious actors exploit them.
Google’s reported results are promising, particularly in code optimization, scientific computing and security research. However, the figures largely come from the company’s own deployments and should be considered alongside independent evaluations and real-world results from other organizations.
As adoption expands, the model’s practical value will ultimately depend on its reliability, operating costs, security safeguards and performance across different workloads.
For now, Gemini 4 Argon represents Google’s latest effort to position advanced AI as a practical tool for solving complex technical and business problems, rather than simply generating responses to user prompts.
