📰 10 sources 13 articles · 30 Sep 2026, 23:07 → 16 days ago

Google launches Gemini 4 Argon, restricting access to cybersecurity experts

Google announced its new flagship model Gemini 4 Argon, describing it as its most advanced model yet, designed for software engineering, cyber defense and enterprise work. The company decided to delay public availability and limit the first phase to a trusted group of cybersecurity experts through the Fairwind program, while granting the U.S. government early access and collecting tester feedback before any wider release.

📍 Where things stand: The model is currently available to a limited group of trusted cybersecurity partners via the Fairwind program, with early access for the U.S. government, and no date has been announced for public availability.

⏱️ Timeline

30 Sep 2026, 23:07 Google began a phased rollout of the Gemini 4 Argon model to trusted cybersecurity partners, evaluating its safety with the U.S. government ahead of a wider release. The model set a record in real-world software engineering, tied with GPT-6 Astra and Grok 4.7 in cybersecurity tests, and topped the Vals index. Google is using it to improve memory efficiency in data centers and intends to strengthen safeguards before making it available to the public.
1 Oct 2026, 00:11 Google announced it will delay making its most powerful AI model available to the public and will release Gemini 4 Argon only to a select group of cybersecurity experts to prevent misuse. Chief AI engineer Koray Kavukcuoglu said in a blog post that safely releasing advanced capabilities at this level requires an incremental approach. Google said it is granting the U.S. government early access to the model and collecting tester feedback before a wider release.
1 Oct 2026, 00:38 Google said Gemini 4 Argon is designed to refuse requests that could help carry out cyberattacks or develop chemical, biological or nuclear weapons, and that it will grant the U.S. government early access and collect tester feedback before a wider release.
1 Oct 2026, 02:20 Google announced that Gemini 4 Argon outperforms its competitors in 12 of 18 tests, scoring 77.9% on the DeepSWE software engineering benchmark versus 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra, and recording only 0.7% on the Gray Swan indirect prompt injection test, the lowest among the compared models.
1 Oct 2026, 02:53 GOOGL shares rose 1% after the close of trading following Google's unveiling of Gemini 4 Argon. The system is initially released to a limited group of trusted cybersecurity experts through the Fairwind program, to be followed by a broader commercial launch for enterprise users, developers and subscribers.
1 Oct 2026, 05:27 Google said the model can detect, verify and fix software vulnerabilities, in addition to handling coding, research and visual tasks, and that it outperforms models from OpenAI and Anthropic in some benchmarks amid increasing competition in the field.
1 Oct 2026, 10:06 Google said the release follows an incremental approach similar to what Anthropic did with its advanced model, granting the U.S. government early access to the model and collecting tester feedback before making it widely available.
1 Oct 2026, 11:16 Google announced that Gemini 4 Argon is released first to a select group of trusted cyber defenders via the Fairwind program, at an introductory price of two dollars per million input tokens and ten dollars per million output tokens, saying the model is designed for deep reasoning tasks in software engineering and enterprise fields such as law, finance and cyber defense, and that it is already used internally.
1 Oct 2026, 11:57 Google announced the launch of Gemini 4 Argon, the flagship model of the Gemini 4 generation, on September 30, 2026, aiming to close the competitive gap with Anthropic and OpenAI. The initial release focuses on the cyber defense sector through the Fairwind program, with support for up to one million tokens, scoring 77.9% on the DeepSWE v1.1 benchmark and 68% on CWE-bench v1; no date was announced for public availability.
1 Oct 2026, 18:50 Alphabet officially announced the Gemini 4 Argon model as its most advanced AI model to date, featuring an expanded output window of up to one million tokens that enables deep reasoning on multi-step tasks. It excels in software engineering, legal drafting and financial analysis, and includes cyber defense capabilities that allow it to independently detect, verify and patch software vulnerabilities.

Timeline written by AI from the sources below; updated as new reports arrive.

How the sources covered it

Coverage by tokenpost.com, ETV Bharat and Cybersecurity focused on technical and commercial details of the cybersecurity model such as benchmarks, pricing and the launch date, while Yahoo Tech, The Hindu, Firstpost and others led with the restriction of access, the delay of public availability and misuse concerns, some quoting Google engineer Koray Kavukcuoglu and the U.S. government's early access.

tokenpost.com Led with the phased rollout to trusted cybersecurity partners and the safety evaluation with the U.S. government, highlighting benchmark records, topping the Vals index and use of the model to improve memory efficiency in data centers.
Yahoo Tech Across its several items, focused on delaying public availability and limiting it to cybersecurity experts, quoted Koray Kavukcuoglu saying safe release requires an incremental approach, and detailed performance benchmarks, GOOGL's 1% after-hours rise and the Fairwind program.
The Hindu Led with the delay of public availability and the initial release limited to an audited group of cybersecurity experts, emphasizing that the model is designed to refuse requests related to cyberattacks or developing chemical, biological or nuclear weapons, and noting early U.S. government access.
Firstpost Opened with the cybersecurity focus and the model's availability to a limited group via the Fairwind program, highlighting its ability to detect, verify and fix software vulnerabilities alongside coding, research and visual tasks, and its outperforming OpenAI and Anthropic models on some benchmarks.
ETV Bharat Highlighted the model's technical and enterprise nature and deep reasoning in software engineering, law, finance and cyber defense, and included pricing details of two dollars per million input tokens and ten dollars per million output tokens and its internal use.
Cybersecurity Set the launch date at September 30, 2026, linked it to competition with Anthropic and OpenAI, cited one-million-token support and results of 77.9% on DeepSWE v1.1 and 68% on CWE-bench v1, noting that no date was announced for public availability.