Artificial Intelligence
Google launches Gemini 4 Argon amid internal doubts
Google released Gemini 4 Argon to selected cyber defenders first, said it improved several coding and security benchmarks, and disputed reports that employees thought it worked poorly on some real-world tasks.
- Google unveiled Gemini 4 Argon, which it said was the first new version of its main model line since Gemini 3 in November.
- Google said the first users would be trusted cyber defenders through its Fairwind Program, with customers on its paid API and Google AI Ultra plans to follow.
- Google said Argon scored 77.9% on DeepSWE v1.1 and 68% on CWE-bench, and that it raised the output limit to 1 million tokens from 64,000.
- Google said Wiz was already using Argon in its Scan for Good initiative and that the model found a critical flaw in healthcare software used by hospitals worldwide.
- Bloomberg reported that some Google employees thought the model performed well on tests but less well in everyday use, and Google told Bloomberg that saying it underperformed in coding would be inaccurate.
Google launched Gemini 4 Argon, its latest flagship model, and said it would first reach a small group of cyber defenders before wider release. Koray Kavukcuoglu, who runs Google DeepMind day to day, introduced it in the announcement.
Google said access would come through its Fairwind Program for trusted defenders and that the US government was also involved in a voluntary pre-release access process. It said those early users would get the model without its cyber restrictions.
The company said Wiz was already using Argon through its Scan for Good initiative. Google said the model found a critical vulnerability in software used by hospitals around the world, which it said earlier frontier models had missed.
Google said Argon reached 77.9% on DeepSWE v1.1 and 68% on CWE-bench. It also increased the output limit to 1 million tokens and set an introductory price of $2 per million input tokens and $10 per million output tokens.
Bloomberg reported that some employees were sceptical, saying the model did well on benchmarks but less well in practical use and struggled with some coding tasks. Google told Bloomberg that it would be wrong to say the model underperformed in coding, while one employee told Bloomberg there was broad agreement that it was at the frontier.
Named in this story
People
- Koray Kavukcuoglu
- described the model in Google’s announcement and runs Google DeepMind day to day
Companies
- unveiled Gemini 4 Argon and set the rollout terms
- Wiz
- used the model through its Scan for Good initiative
Organisations
- Google DeepMind
- the Google unit associated with the launch and model products
- Fairwind Program
- Google’s access route for trusted cyber defenders
- Scan for Good initiative
- Wiz programme that used the model to look for flaws
- Bloomberg
- reported employee doubts and Google’s response to them
Governments and agencies
- US government
- took part in Google’s voluntary pre-release access process
Products and systems
- Gemini 4 Argon
- the new Google model at the centre of the release
- DeepSWE v1.1
- the software engineering benchmark where Google said the model scored 77.9%
- CWE-bench
- the security benchmark where Google said the model tied for first
- Google AI Ultra
- the subscription tier whose users were next in line for access
How the source tells it
The article was mostly straight reporting, with Google’s launch language and benchmark claims set against a short section on employee scepticism.
- vendor boosterism Google’s own promotional wording and benchmark-lead claims are repeated in attributed form
- scepticism anonymous employees and Google’s denial are used to frame a dispute over how the model worked in practice