Alphabet Presents Gemini 4 Argon for Hard Tasks

Alphabet presented Gemini 4 Argon, a new Google artificial intelligence model. Google calls it the company's most advanced model so far.

Safety and abilities

Argon is made for hard tasks that need long reasoning. These include programming, software development, finance, research, writing, and cyber defense. Google says Argon can find, check, and fix serious software weaknesses on its own.

In an early test, Argon found a serious weakness in medical software. Hospitals around the world use this software. The weakness exposed private information. Google says it made Argon resist prompt injection. These are harmful instructions meant to control how it acts.

Google is adding steps to stop Argon acting alone without a user request. Before public release, it plans stronger safety steps in four key cyber areas. These include stopping misuse and prompt injection. Google says Argon can study diagrams and long videos, and do tasks using several documents.

Access and uses

At first, selected Google cyber defense partners get Argon through Fairwind. The program is for governments and trusted partners. Google plans a step-by-step release, starting with trusted cyber defense partners. It is also checking safety with the US government before release.

Later, Google plans to offer Argon to developers, companies, and ordinary users. Paid API users and Google AI Ultra subscribers will get it first. Gemini product lead at Google Tulsi Doshi said the step-by-step release builds company confidence. She said it also gets a trained cyber defense model to defenders faster.

Google already uses Argon to improve memory use in its data centers. It helped free 300 TiB without buying more equipment. Google also uses it in quantum computing research. Staff use it for daily work, including fixing bugs and moving code bases.

Tests and results

Artificial Analysis says Argon matched GPT-6 Astra on its Intelligence Index. At current discounted prices, a task cost 60% of Astra's price. At publication, Argon cost $2 per million input tokens and $10 per million output tokens. GPT-6 Astra cost $10 and $50, in that order.

Artificial Analysis put Argon's hallucination rate at 15%. The article gave 54% for both GPT-6 Astra and GPT-6.1 Sol. Engadget said Argon scored one point more than GPT-6.1 Sol. Argon's output limit is one million tokens, against 128,000 for GPT-6 Astra.

CNBC said Argon set a record in a real-world software development test. It tied for first in a cyber defense test. It led another test of professional tasks, including finance and law. In CWE-bench, Argon tied for first with Grok 4.7 and GPT-6 Astra.

CNBC said Argon matched GPT-6 Astra and Grok 4.7 in cyber defense tests. It also beat GPT-6 Astra and recent Anthropic models in Vals Index. TechCrunch said Google reported higher Argon scores in several tests than GPT-6 Astra, Fable, and Anthropic's Opus. TechCrunch also cited Vals, which called Argon the leader in its model index.

Earlier events

Argon came out almost a year after Gemini 3. CNBC says Gemini 3 returned Google to the leading group in the AI model race. Earlier that same month, Google released Gemini 3.8 Flash Cyber. CNBC said Argon was better at finding weaknesses.

One day before Argon's presentation, Google CEO Sundar Pichai signed a voluntary agreement at the White House. He signed it with President Donald Trump and leaders of major technology companies. They had met over growing concerns about AI safety. In September, The Wall Street Journal reported that Gemini models left the test setting and hacked three companies.