Google has made its newest flagship AI model, Gemini 4 Argon, available worldwide, positioning it as a system built for extended, multi-step reasoning across complex professional tasks.

Developers and enterprises can access the model through the Gemini API in AI Studio and the Gemini platform. Google AI Ultra subscribers can also use it in the Antigravity agentic environment, as well as in the Gemini app and AI Mode in Search.

According to Google, Argon ranks first on the Vals Index, a benchmark that measures model performance across economically critical sectors including finance, coding, legal, and tax work. Cross-sector rankings of this kind have become a standard way for AI labs to argue that their models are ready for regulated, document-heavy professional workflows.

Launch pricing is set at $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95% — a relevant detail for teams that repeatedly process the same long documents. Once the introductory period ends on November 30, 2026, rates will double to $4 and $20 per million tokens, respectively, giving teams a fixed evaluation window at launch economics before costs step up. Through the API, the model supports an industry-leading unified input-output context window of 1 million tokens, large enough to hold extensive codebases or document collections in a single pass rather than breaking work into smaller chunks.

Google also shared early internal results that it says point to meaningful productivity gains. Running Argon inside Antigravity increased agentic code submissions by more than 35% while cutting code rollbacks by nearly a third. The Chrome team used the model to untangle a roughly 6-million-line legacy folder within Chromium's 40-million-line codebase, accelerating an architectural refactor that would otherwise have taken months. Optimizations contributed to the open-source ADK framework reduced in-memory session preparation time by 40% and improved cold imports by 21%, while Google's Gemma models achieved a 2.85x inference speedup on A100 hardware with full token parity. Because these figures come from Google's own engineering teams rather than independent measurement, they show how the model performed on large-scale codebases under conditions the company controls.

— Google AI Studio (@GoogleAIStudio) October 13, 2026

Real-World Deployments Across Science and Cybersecurity

Beyond software engineering, Argon is being applied to high-stakes scientific and security domains, a set of deployments that extends the model's track record beyond benchmark scores. In quantum computing, the model translated Google's AlphaQubit error-correction decoder into a single optimized hardware kernel, nearly doubling inference throughput. It also produced a formally verified implementation of the arcsin subroutine that cuts spacetime costs by more than 100x compared with a published baseline.

Working with University of Maryland researchers, Google used Argon to identify new counterexamples to Grad's conjecture, a 1967 hypothesis concerning stellarator magnetic field optimization that is central to fusion energy research.

In the life sciences, researchers at Gamow Labs used Argon to detect a chromosomal rearrangement in a newborn with a rare lung disorder that standard software had missed for years. Google's own team, meanwhile, designed five novel serine esterase enzymes within hours, all of which showed confirmed catalytic activity in laboratory testing.

On the security front, Wiz employed Argon to remediate vulnerabilities in government and healthcare infrastructure, including a file download bypass that exposed administrative access to patient records. Google is also expanding its Fairwind Program to hundreds of cybersecurity partners and offering 10,000 researchers six months of free access to Antigravity's highest usage tier, putting the model in the hands of outside organizations for extended, real-world evaluation.

On safety, Google reports that external biological red teaming unlocked fewer than 1% of the bio sub-tasks tested, that shortcut attempts on coding benchmarks fell to 2%, and that Argon leads the Gray Swan Indirect Prompt Injection benchmark with a 0.7% attack success rate across 15 adaptive attempts — a class of attack in which malicious instructions hidden in processed content attempt to hijack a model's behavior.

This report was first published by Metaverse Post.