Artificial Intelligence
Google rolls out Gemini 4 Argon worldwide
Google made Gemini 4 Argon available through several products and tiers, said it outperformed some internal workflows, and paired the launch with pricing, benchmark and safety claims.
- Google rolled out Gemini 4 Argon worldwide and said it was available through AI Studio, the Gemini platform, Google Antigravity, the Gemini app and Google Search depending on subscription.
- Google said free Gemini users and Plus subscribers were limited to Flash models under a new tier structure, while Argon replaced the Gemini Pro class and more Gemini 4 models would follow.
- Google priced Argon at $2 per million input tokens and $10 per million output tokens at launch, with cached inputs discounted by 95 per cent.
- Google said Argon topped the Vals Index but still trailed Anthropic's Claude Opus 5.5 in independent rankings.
- The company said its own use of Argon lifted agent-driven code submissions by more than 35 per cent and cut code rollbacks by nearly a third.
Google rolled out Gemini 4 Argon worldwide and said the model was available through AI Studio, the Gemini platform, Google Antigravity, the Gemini app and Google Search, with access varying by subscription. Free users and Plus subscribers were moved onto Flash models under a new tier structure, while Argon took over from the Gemini Pro class.
Google said the launch price was $2 per million input tokens and $10 per million output tokens, with cached inputs charged at a 95 per cent discount. It said Argon led the Vals Index but still lagged Anthropic's Claude Opus 5.5 in independent overall rankings, and the article said it also sat behind Meta's Muse Wonder Full because Google shipped later.
The company said internal use of Argon increased agent-driven code submissions by more than 35 per cent and reduced code rollbacks by nearly a third. In Chromium, Google said the model helped complete a refactor across hundreds of files in a 40-million-line codebase in a few weeks rather than months.
Google said Argon also worked overnight with the Agent Development Kit, improving session preparation time by 40 per cent and speeding up cold imports by 21 per cent. It said the model could generate code for 3D scenes, a Python raytracing demo and autonomous Blender work through MCP, and the article added that Google was reportedly testing a codenamed successor, Carbon.
Google said it was giving 10,000 researchers six months of free access to the highest Antigravity tier and was opening a separate Fairwind programme for cybersecurity defenders. It also said Argon had its strongest safety results so far, with lower rates on several tests, while the article noted that Google did not say whether Gray Swan had used the same test setup as before.
Named in this story
People
- Alexandr Wang
- was Meta's AI chief and was cited on 3D model training
Companies
- rolled out Gemini 4 Argon and supplied the launch claims
- Anthropic
- was the company behind Claude Opus 5.5 in the benchmark comparison
- Meta
- was the company behind Muse Wonder Full in the benchmark comparison
Products and systems
- Gemini 4 Argon
- was Google's new flagship model
- AI Studio
- was one route through which Argon was made available
- Gemini platform
- was another access point for Argon
- Google Antigravity
- gave AI Ultra subscribers access to Argon
- Google AI Ultra
- was the subscription tier that included Antigravity access
- Gemini app
- gave AI Ultra and Pro subscribers access to Argon
- Google Search
- included AI mode access to Argon
- Gemini Pro class
- was replaced by Argon as Google's frontier series
- Claude Opus 5.5
- was the independent benchmark rival that outranked Argon
- Muse Wonder Full
- was the other benchmark rival that moved ahead after a slower launch
- Carbon
- was the codenamed successor Google was reportedly testing
- Chromium
- was the codebase where Google said Argon helped complete a refactor
- Agent Development Kit
- was the open-source toolkit Argon used overnight
- Gemma models
- were the open models whose inference speed Google said improved
- vLLM
- was the inference engine used in the speed test
- Blender
- was controlled by Argon in the 3D demo
- MCP
- was the interface Argon used in the autonomous demo
- Fairwind program
- gave cybersecurity defenders access to Argon without cyber safeguards
How the source tells it
The article read like a launch brief, mixing promotional language, self-reported gains and selective benchmark and safety claims.
- boosterism launch wording and benchmark ranking claims were used to present the model as a flagship release
- triumph self-reported productivity gains and work finished in weeks were used to frame the rollout as a success
- reassurance safety percentages and strongest-to-date language were used to calm concerns
- speculation hedged reporting on a codenamed successor and likely future variants was presented as an expectation