Google unveils Gemini 4 Argon with leading prompt-injection test score

Google unveiled Gemini 4 Argon on Wednesday, calling it its most capable model for coding, office work and cyber defense. Argon recorded a 0.7% attack success rate on Gray Swan's indirect prompt-injection benchmark, which tests malicious instructions hidden in content an AI reads across 15 attempts. Lower scores indicate fewer successful attacks. Claude Opus 5.5 and Claude Fable 5.1 each scored 1.0%. GPT-6 Astra scored 8.5%, while Grok 4.6 scored 51.8% and Kimi K3 scored 52.7%. On DeepSWE v1.1, a test of long software engineering tasks, Argon scored 77.9%. Claude Opus 5.5 scored 74.2%, GPT-6 Astra scored 74.1% and Claude Fable 5.1 scored 67.4%. Gemini 3.6 Flash scored 49% on the same test in July. Google calculated Argon's DeepSWE score itself, while the rival scores came from a public leaderboard and company reports. Google's comparison shows Argon leading on 12 of 18 benchmarks, tying one and trailing on five. Argon can produce up to 1 million tokens in one reply, up from 64,000 tokens. Google is first providing Argon to vetted security teams through its limited-access Fairwind Program without cyber guardrails that would normally restrict hacking assistance. Fairwind launched September 2 with more than 650 partners, including governments and critical infrastructure operators. Google says defenders need such capabilities to find security flaws before criminals do. Google says the same capabilities could help attackers, making a phased rollout the only safe path. Google is also participating in the U.S. government's voluntary process for pre-release model access. Google says wider access will begin as soon as possible with paid API customers and Google AI Ultra subscribers. On Google's internal Wiz Penetration Test Benchmark, Argon solved 70.9% of web-application exploit tasks on the first try, compared with 58.2% for Gemini 3.8 Flash Cyber. Google launched Gemini 3.8 Flash Cyber with Fairwind as a restricted model. Google says Argon helped security firm Wiz find a critical flaw in healthcare software used by hospitals worldwide that earlier frontier models had missed. An early version of Anthropic's Claude Mythos helped identify 271 Firefox vulnerabilities that Mozilla subsequently patched. OpenAI has a similar restricted-access initiative called Trusted Access for Cyber. Google released smaller Flash models in July but did not release the promised Gemini 3.5 Pro. Alphabet shares fell about 4.4% that month. Argon's launch came the same day President Trump unveiled a voluntary, penalty-free AI accord signed by Google's leadership. Introductory API pricing is $2 per million input tokens and $10 per million output tokens. Google has not said when introductory pricing ends; standard rates are $4 and $20 per million input and output tokens, respectively.