Google launches Gemini 4 Argon globally at $2 per million input tokens

Google's Gemini 4 Argon became globally available today at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95 percent off. The release positions the model as Google's most capable frontier system to date, ranking first on the Vals Index, an aggregate measure of model performance across finance, coding, legal, and tax workloads, and carrying a unified 1 million-token input-output window via the Gemini API.

Google said Argon "excels at deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science." Distribution spans four surfaces: developers via the paid Gemini API tier, AI Studio, and Android Studio; enterprises via Gemini for business, the API, and Google Antigravity; and Google AI Ultra and Pro subscribers via the Gemini app and AI Mode in Search.

What does Argon deliver inside Google?

Internal metrics released by Google show Argon in Antigravity increased agentic code submissions (autonomous code completions produced without human intervention) by more than 35 percent while reducing code rollbacks by nearly one-third across participating software teams.

• Chrome refactor across Chromebooks (Chromium's 40 million-line codebase): Argon separated core webpage handling from a roughly 6 million-line legacy folder, completing an architectural refactor (a structural code reorganization that improves maintainability) across hundreds of files in weeks rather than months.
• Google ADK Python (Google's open-source Agent Development Kit): overnight autonomous work produced live optimizations cutting in-memory session prep time by 40 percent and accelerating cold imports (first-time loading of a module into memory) by 21 percent.
• xTB quantum chemistry package (Semiempirical Extended Tight-Binding): made 2.3x faster at predicting molecular structures and 1.8x faster at conformer search (identifying plausible 3D shapes a molecule can adopt).
• Gemma 4 E2B (an open-weights model with effective 2B parameters): 2.85x end-to-end inference speedup on an Nvidia A100 via vLLM (a high-throughput LLM serving system), pushing throughput from approximately 37,900 to over 108,000 tokens per second with 100 percent token parity and zero measured quality drop.

Google added that the Gemma optimizations, spanning attention tuning and speculative decoding (a technique generating draft tokens ahead of the main model to reduce latency), would normally require months of engineering rather than days.

What does Argon deliver for security and defense?

Wiz deployed Argon under its "Scan for Good" initiative, which delivers cybersecurity services to critical infrastructure at no cost. In collaboration with a federal agency, Argon identified and remediated a gap risking exposure of sensitive data and government infrastructure, according to Google. In healthcare, Argon chained complex vulnerabilities to discover and rapidly fix a public file download bypass that exposed a master key and granted administrator access, safeguarding hundreds of thousands of private patient records.

External bio red teaming unlocked less than 1 percent of bio sub-tasks Google tested. On the Gray Swan Indirect Prompt Injection (IPI) benchmark, which measures resistance to malicious instructions embedded in web pages, emails, or code, Argon cut attack success to 0.7 percent across 15 adaptive attempts.

What scientific results has Argon produced?

In quantum computing, Argon translated an internal version of AlphaQubit, Google's machine-learning quantum error correction decoder (a system that identifies and corrects errors in quantum bits), into a single hardware kernel (a low-level routine executed directly on a processor), nearly doubling inference throughput. An Argon-written arcsin implementation, a building block in recent quantum simulation algorithms, was formally verified in Lean (a programming language and interactive theorem prover) and cuts spacetime cost (combined memory and processing time) by more than 100x versus a published quantum CORDIC baseline, an iterative method for trigonometric computation.

In nuclear fusion, a Google team alongside Professor Matt Landreman of the University of Maryland extended counterexamples to Grad's conjecture, the 1967 hypothesis asserting that exact mathematical solutions for stellarator magnetic fields (3D magnetic configurations confining plasma in fusion reactors) were impossible. The new results give the fusion community broader benchmarks for testing simulation codes used in future stellarator engineering.

In rare disease research, Gamow Labs used Argon on de-identified genomes (datasets with personal identifiers removed) from newborns with rare developmental lung disorders. Argon pinpointed a chromosomal rearrangement that routine software had missed for years. On internal benchmarks, Argon scored 57 percent on RareBench (rare-disease diagnosis assistance) and 52.2 percent on LabBench (wet-lab reasoning).

In enzyme design, Google scientists worked with Argon to engineer serine esterases (enzymes that cleave ester bonds, a class of chemical linkages, on demand) against the benchmark molecule 4MU-acetate, which fluoresces under UV light when cleaved. Argon assembled data, performed model tuning, and reached five novel designs in hours; laboratory testing confirmed robust catalytic activity across those candidates.

To accelerate downstream research, Google will offer 10,000 scientists the highest usage tier in Google Antigravity free of charge for six months. Applications are open, with rolling onboarding subject to eligibility and capacity.

How does Argon perform on physical reasoning?

On the Embodied Reasoning (ERQA) benchmark, which tests a model's ability to solve physical, spatial problems in 3D environments, Argon scored 87 percent using pure reasoning, ahead of all other frontier models Google compared against. In an internal demonstration, Argon watched a raw video of a human playing cornhole, mapped spatial coordinates, calculated required whole-body joint movements, and autonomously generated simulation code. The code was deployed onto Apptronik's Apollo humanoid, which sank seven consecutive shots in a real-world trial.

Google called the resulting Video-to-Sim-to-Real pipeline (the process of converting ordinary video into a robot's training environment) a method that compresses weeks of manual simulation modeling into a few hours, framing it as a scalable, infinite training ground for robotics.

How does Argon handle safety and agent alignment?

Under Google's Frontier Safety Framework (a third-party risk evaluation protocol for advanced capabilities), Argon was tested across misuse, jailbreak, and agentic risk categories. On an adversarial set (curated challenging inputs designed to provoke failures), violations fell nearly two-fold relative to previous Gemini releases. On coding benchmarks where models face no reason to hold back, shortcut attempts dropped to 2 percent, the best result across Google's Gemini line to date.

Argon was trained to respect system boundaries, report negative findings honestly, and pause for human confirmation before irreversible actions. A separate automated misalignment monitor reads the agent's step-by-step reasoning and actions in real time, blocking unintended workarounds across internal deployments, capability evaluations, and reinforcement learning training runs.

Who can use Argon today, and what comes next?

• Developers: build on the paid Gemini API tier; Google AI Ultra subscribers gain access through Antigravity, AI Studio, and Android Studio.
• Enterprises: test in Gemini for business, the Gemini API, and Google Antigravity.
• Subscribers: Google AI Ultra and Pro tier users can access Argon in the Gemini app and AI Mode in Search.
• Scientists and researchers: apply for six months of free highest-tier Antigravity access.
• Cybersecurity defenders: apply to the Fairwind Program (Google's restricted-access initiative for vetted security professionals) for unconstrained Argon access without cyber guardrails.

Argon operates under a 1 million-token unified context limit via the Gemini API, with documentation covering a new continuation protocol (a method for handling inputs or outputs longer than a single request window). Google signaled that Argon is the first release in the broader Gemini 4 family, indicating that subsequent Gemini 4 variants are positioned to follow in the coming quarters.