Google launches Gemini 4 Argon, limits early access to cyber defenders

Google unveiled Gemini 4 Argon on September 30, 2026, describing it as the newest frontier model from Google DeepMind. According to Google's official announcement, Argon is designed for deep reasoning across long, complex workflows in software engineering, cybersecurity defense, and enterprise knowledge work such as legal and financial analysis.

What the model can do

Google says Argon supports an output limit of 1 million tokens, up from 64,000 tokens in earlier models, allowing it to sustain longer multi-step tasks. The company's announcement states the model can autonomously find, validate and patch software vulnerabilities, and that it has strong multimodal and video-understanding capabilities.

On benchmarks cited by Google and reported by 9to5Google, Argon scored 77.9% on DeepSWE v1.1, a real-world software engineering test, ahead of Claude Opus 5.5 (74.2%) and GPT-6 Astra (74.1%). On CWE-bench v1, a measure of vulnerability remediation, Google said Argon tied for first place with a score of 68%. Google also reported Argon ranked first on AutomationBench, a test of business-process automation, with a score of 51.3%, and scored 91.7% on LVBench, a long-video-understanding benchmark.

Limited release through the Fairwind Program

Rather than a broad public launch, Google is initially rolling out Argon to what it calls "trusted cyber defenders" through its Fairwind Program, according to the company's blog post and a report from The Hacker News. Google said it plans to release a version of Argon without cyber guardrails specifically to trusted defenders and internal teams so they can use its full capabilities, while continuing to strengthen safeguards before any wider release. 9to5Google reported the model is also set to roll out soon to Google AI Ultra subscribers and paid API customers.

Google said it is deploying what it calls misalignment mitigations that monitor the model's chain-of-thought and actions and can stop execution when necessary, according to The Hacker News. Google also said Argon topped Gray Swan's IPI benchmark, which tests resistance to indirect prompt-injection attacks.

Pricing

Google set introductory pricing for Argon at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens, according to the company's announcement. Standard pricing after the introductory period rises to $4 per million input tokens and $20 per million output tokens.

Market reaction and internal debate

Alphabet shares moved higher following the Gemini 4 Argon announcement, according to a Yahoo Finance report. The same report, citing Simply Wall St analysis, said Alphabet closed at $344.08 on October 1, 2026, up 9.2% year-to-date and 40.4% over the prior year, though down about 4.4% over the previous 90 days. Yahoo Finance noted a valuation disconnect between different models: one approach put Alphabet's fair value at $202.62, implying the stock was about 70% overvalued, while a price-to-earnings-based comparison (17.2x versus a 27.3x fair ratio) suggested the opposite.

Android Headlines, citing Bloomberg reporting, said some Google employees privately raised concerns that while Argon performs strongly on standardized benchmarks, it struggles on some real-world coding and front-end design tasks. Edwin Chen of Surge AI was quoted comparing the pattern to "a student getting top SAT scores without developing actual real-world skills," a critique sometimes called industry "benchmaxxing."

Koray Kavukcuoglu of Google DeepMind said, as reported by Android Headlines, that it is "a certainty Google will stay at the AI frontier," defending the model's capabilities amid the internal debate.

Background

The Argon launch follows what Android Headlines described as pressure on Google after an earlier Gemini 3.5 Pro model was scrapped and amid reported departures of senior AI talent from the company. Android Headlines reported that Google's own materials claim Argon leads on measures related to safety, cybersecurity, conversation quality and video metadata extraction.