Google has unveiled Gemini 4 Argon, its new frontier artificial-intelligence model, presenting it as a system built for software engineering, office tasks and cyber defense. The launch places Argon in direct competition with the latest models from Anthropic and OpenAI, while Google is limiting its initial cybersecurity access because the model can operate without the refusals that typically restrict assistance with hacking-related work.
According to Decrypt, Argon scored 77.9% on DeepSWE v1.1, a benchmark designed to measure whether an AI can complete lengthy and complicated software-engineering assignments. That result was higher than the scores reported for Claude Opus 5.5, at 74.2%, and GPT-6 Astra, at 74.1%. Claude Fable 5.1 scored 67.4%.
Google’s own comparison table showed Argon leading 12 of 18 benchmarks, tying on one and trailing on five. The tests covered areas including coding, science and computer control. However, the comparisons should be read carefully: Google calculated Argon’s DeepSWE result, while the figures for competing models came from a public leaderboard and company reports.
Argon also expands the amount of material it can process in a single response. Its maximum output has risen to 1 million tokens, compared with 64,000 for the previous figure cited in the launch material. A token is a portion of text, roughly three-quarters of a word, meaning the larger limit corresponds to approximately 750,000 words rather than about 48,000.
Cybersecurity performance and risks
The most prominent part of the launch is Argon’s performance on tests involving indirect prompt injection. In this type of attack, a malicious instruction is hidden inside material an AI agent is asked to read, such as an email. The goal is to persuade the system to follow the embedded instruction instead of the user’s request.
On Gray Swan’s Indirect Prompt Injection benchmark, Argon recorded an attack success rate of 0.7%. The test measures how often an attack works within 15 attempts, so a lower score indicates stronger resistance. Claude Opus 5.5 and Claude Fable 5.1 each recorded 1.0%. GPT-6 Astra scored 8.5%, while Grok 4.6 and Kimi K3 were tricked in 51.8% and 52.7% of cases, respectively.
Google is initially making Argon available to vetted cybersecurity teams through the Fairwind Program, its restricted-access cyber-defense initiative. The program launched on September 2 with more than 650 partners, including government bodies and critical-infrastructure operators.
Argon will be provided through the program without cyber guardrails, meaning the built-in refusals that generally prevent AI models from assisting with hacking tasks. Google’s stated rationale is that defenders need systems capable of reasoning like attackers so they can identify and fix weaknesses before criminals exploit them. The same capability can create risks, however, which is why Google is using a phased release and participating in the U.S. government’s voluntary process for pre-release model access.
Google is not alone in restricting powerful cyber tools. Anthropic previously used an early version of Claude Mythos to help identify 271 vulnerabilities in Firefox, which Mozilla subsequently patched. OpenAI has also created its Trusted Access for Cyber program.
Rollout, pricing and market picture
Argon’s results also exceed those of Gemini 3.8 Flash Cyber, the restricted Google model introduced with Fairwind. On Google’s Wiz Penetration Test Benchmark, which asks an AI to produce working exploits for real web-application flaws without access to the underlying code, Argon solved 70.9% of cases on the first attempt, compared with 58.2% for Gemini 3.8 Flash Cyber.
Google said Argon also helped cybersecurity company Wiz identify a critical flaw in healthcare software used by hospitals worldwide. Earlier frontier models had not found the vulnerability, according to Google.
The broader release is expected to begin with paid API customers and Google AI Ultra subscribers as soon as possible. Introductory pricing is set at $2 per million input tokens and $10 per million output tokens. Google said standard rates will be $4 and $20, respectively, but has not specified when the introductory period will end.
The launch follows a difficult period for Google’s AI product timetable. In July, the company released smaller Flash models but did not deliver the promised Gemini 3.5 Pro, and Alphabet shares fell about 4.4%. Argon was also announced on the day President Donald Trump introduced a voluntary, penalty-free AI accord that Google’s leadership signed.
Moneyiar market data at 09:23-09:24 on October 1, 2026, showed mixed moves among the listed digital assets:
- Ethereum Name Service (ENS) was $6.9205686646, down 0.9% over 24 hours but up 4.24% over seven days.
- Decimal (DEL) was $0.4320905509, up 2.79% in 24 hours and down 9.09% over seven days.
- Loaded Lions (LION) was $0.0015460096, up 2.1% over 24 hours and 4.66% over seven days.
- Railgun (RAIL) was $2.3902096364, down 4.1% over 24 hours and 13.71% over seven days.
