Artificial Intelligence · Cybersecurity
Google launches Gemini 4 Argon to cyber defenders
Google introduced Gemini 4 Argon, opening it first to selected cybersecurity defenders before wider customers, and said the model topped or matched several coding benchmarks while some employees remained doubtful.
- Google unveiled Gemini 4 Argon on Wednesday as the first new generation of its main model since Gemini 3 last November.
- Google said selected cyber defenders would get first access through its Fairwind Program and the United States government’s voluntary pre-release access process, before paying API users and Google AI Ultra subscribers.
- Google said Wiz was already using Argon in its Scan for Good initiative, and that the model found a critical vulnerability in healthcare software used by hospitals worldwide.
- Google said Argon scored 77.9% on DeepSWE v1.1 and 68% on CWE-bench, and Axios reported that it outperformed OpenAI’s GPT-6 Astra on several coding and knowledge-work tests.
- Google said Argon’s output limit would rise to 1 million tokens and its introductory price would be $2 per million input tokens and $10 per million output tokens.
Google unveiled Gemini 4 Argon on Wednesday and described it as the first new version of its main model since Gemini 3 last November. It said the model would reach a small group of cybersecurity defenders before paying customers and Google AI Ultra subscribers.
Google said the first users would come through its Fairwind Program for trusted cyber defenders and through the United States government’s voluntary pre-release access process. It also said those defenders would use Argon without the cyber guardrails that would apply more widely.
The company said Wiz was already using Argon in its Scan for Good initiative, and that the model found a serious flaw in healthcare software used by hospitals worldwide. Google also said Argon reached 77.9% on DeepSWE v1.1 and 68% on CWE-bench.
Google said the model would allow outputs up to 1 million tokens and would be priced at $2 per million input tokens and $10 per million output tokens at launch. Axios reported that Google said Argon beat OpenAI’s GPT-6 Astra on several coding and knowledge-work benchmarks.
Bloomberg reported that some Google staff doubted how well the model worked outside benchmarks and said it struggled with some coding tasks. Google told Bloomberg that it would be wrong to say Gemini 4 underperformed in coding.
Named in this story
People
- Koray Kavukcuoglu
- said Argon marked a new phase for frontier intelligence
- Tulsee Doshi
- head of Gemini products at Google DeepMind and spoke to Axios
Companies
- unveiled Gemini 4 Argon and set the rollout terms
- Wiz
- was already using Argon in its Scan for Good initiative
- OpenAI
- owned the GPT-6 Astra model used in comparison claims
Organisations
- Google DeepMind
- the unit now run day to day by Kavukcuoglu
- Fairwind Program
- the access route for trusted cyber defenders
- Scan for Good
- the initiative through which Wiz used Argon
- Axios
- reported the benchmark comparisons and carried Doshi’s comments
- Bloomberg
- reported internal scepticism and Google’s denial of underperformance
Governments and agencies
- United States government
- ran a voluntary pre-release model access process alongside Google
Products and systems
- Gemini 4 Argon
- the model being launched
- Gemini 3
- the previous flagship generation named as the comparison point
- Google AI Ultra
- the subscription tier next in line for access
- DeepSWE v1.1
- the software-engineering benchmark Argon scored highly on
- CWE-bench
- the security-flaw benchmark on which Argon tied for first
- GPT-6 Astra
- the OpenAI model Google said Argon beat on several benchmarks
How the source tells it
The piece mixed Google’s launch messaging and benchmark claims with a short note of internal scepticism.
- hype Google’s self-promotional framing and benchmark-led comparisons were repeated without much outside checking
- scepticism Bloomberg-reported employee doubts about real-world performance were included as a counterpoint