Google launches Gemini 4 Argon for complex tasks

Google has launched Gemini 4 Argon, a model designed to handle long and difficult tasks in programming, finance, legal work, cybersecurity, and science. It is now available globally to developers, businesses, and some subscribers to Google’s AI services.

Argon is not positioned as a chatbot for answering quick questions. Its goal is to work for longer, break problems into multiple steps, and carry out complex actions with less human supervision.

What Gemini 4 Argon can do

Google says Argon ranks first on the Vals Index, an evaluation that measures model performance in areas such as programming, finance, legal work, and tax. The company also says it outperforms other advanced models in various tests, although those results come from evaluations reported by Google itself.

The model can analyze lengthy documents, review codebases, study data over long periods, and build interactive tools. In practice, that could mean reviewing an entire contract, locating errors in an enterprise application, or comparing thousands of financial records without splitting the work across dozens of queries.

Argon includes a 1 million token context window, a measure of text capacity that allows it to process large amounts of information in a single session. Google has also improved its ability to follow style instructions, respond more directly, and avoid unnecessary repetition.

Results highlighted by Google

The company says its teams are already using Argon in internal projects with measurable results:

• In Antigravity, Google’s tool for developing software with AI agents, teams increased agent-generated code submissions by more than 35% and reduced reversions by nearly one-third.
• In the Chromium project, Argon helped reorganize code related to browser features within a codebase of about 40 million lines. Google says work expected to take months was completed in weeks.
• In Google’s agent development kit, the model cut preparation time for long sessions by 40% and accelerated cold imports by 21%.
• In the xTB quantum chemistry package, optimizations achieved speeds of up to 2.3 times faster when predicting molecular structures.
• In open Gemma models running on an A100 GPU, Google recorded a 2.85 times speed improvement, with no loss of quality or changes to the generated tokens.

These figures describe specific Google tests and projects. They do not guarantee that Argon will deliver the same results for every company or codebase.

It also targets science, robotics, and cybersecurity

Google is using Argon to analyze problems in quantum computing, nuclear fusion, enzyme design, and rare diseases. In one cited case, the model helped identify a chromosomal alteration that routine software had failed to detect for years. Gamow Labs evaluations gave it a 57% score for assistance with rare disease diagnoses and 52.2% for laboratory reasoning.

In robotics, Argon scored 87% on ERQA, a test that evaluates the understanding of physical and spatial problems in three-dimensional environments. In an internal demonstration, it watched a video of someone playing cornhole, calculated the required movements, and generated a simulation that was later transferred to the Apollo humanoid robot. The robot made seven consecutive shots in a real-world test.

Cybersecurity is another major use case. Google says Argon helped find vulnerabilities in government and healthcare infrastructure. In one case, the model helped detect a chain of flaws that allowed private files to be downloaded, a master key to be exposed, and administrator access to be obtained.

Pricing and availability

The launch starts at $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount, which can lower the cost of repetitive tasks.

Argon is available through these channels:

• Developers and businesses can use it through the Gemini API and AI Studio.
• Google AI Ultra subscribers can try it in Google Antigravity and Android Studio.
• Google AI Ultra and Pro users have access through the Gemini app and AI Mode in Search.
• Researchers can apply for six months of free access to the highest usage tier in Antigravity, subject to requirements, capacity, and selection.
• Cybersecurity specialists can request access through the Fairwind Program.

More capability also means more controls

Google says it has added defenses against jailbreaks, malicious instructions, and unauthorized behavior. In its tests, indirect prompt injection attacks, which try to hijack an agent through a website, email, or file, had a 0.7% success rate across a set of 15 adaptive attempts.

The company also says attempts to take unauthorized shortcuts fell to 2% in certain programming tests. Argon can also stop and ask for confirmation before performing irreversible actions. It includes a system that monitors its steps and blocks potential deviations.

For you, the news means Gemini is starting to compete less on answering quickly and more on completing entire jobs: modifying software, analyzing documents, investigating a problem, or controlling a chain of tools. The point to watch is whether these improvements hold up outside Google’s demonstrations, and whether the cost, supervision, and reliability make them practical for tasks where a mistake can have real consequences.