Google Launches Gemini 4 Argon, Its Most Advanced AI Model Yet

Alphabet unveiled Gemini 4 Argon on Wednesday, September 30, 2026, calling it its most advanced AI model so far. Google’s own headline for the launch is “next era of frontier intelligence.” The model is built for coding, professional knowledge work, and cyber defense. 

There is a catch. Argon is rolling out first to a set of trusted cyber defenders through Google’s Fairwind Program, so most people can’t use it yet.

Argon Replaces Gemini 

Google announced Argon on September 30. It follows the cancellation of Gemini 3.5 Pro, which Google first teased at I/O in May. Argon is essentially its replacement.

The launch comes nearly a year after Gemini 3 put Google back at the front of the AI race.

Key Details

• Output limit: Argon can generate up to 1 million tokens in a single output, up from 64,000 before. That helps with long, complex jobs.
• Cyber defense: Google says Argon can find, validate and patch serious software vulnerabilities on its own.
• Security score: On CWE-bench v1, which tests fixing security flaws, Argon ties for first place with 68%.
• Video: On LVBench, a long-video understanding test, Argon scored 91.7%, which Google calls state of the art.
• Safety: Google says Argon leads on the Gray Swan benchmark for resisting prompt injection attacks.

Google’s AI Push 

Argon is Google’s strongest attempt yet to win back the top spot in AI. Google had been overtaken by OpenAI’s GPT-6 Astra and Anthropic’s Fable 5.

Google is also already using the model itself. Internally, it helped free up hundreds of terabytes of memory in Google’s data centers without new hardware. Thousands of Google employees have reportedly used it for coding, research and writing.

Read Also: OpenAI Launches Dots: Always-On AI Agents Powered by GPT-6 Astra

Features and Benchmarks

The numbers below are published by Google, not tested independently. Google lists Argon ahead of GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on these knowledge-work tests.

Argon also leads on Harvey’s legal benchmark, which covers legal research and drafting. The picture is less clear in coding. The New Stack says Argon’s coding scores are mixed, and in some areas it trails rivals by as much as 10 points.

Prices

Google DeepMind’s Logan Kilpatrick said Argon costs $2 per million input tokens and $10 per million output tokens during introductory pricing. The New Stack notes that its output rate is half of what Anthropic charges for Opus 5.5, at $20 per million.

No free public Argon tier has been announced. The word “introductory” suggests the rates could change later.

Availability

Access is limited right now.

• Today: Select cyber partners have it.
• Next: Google says it is “rolling out soon” to Google AI Ultra subscribers and paid API customers.
• Longer term: Argon is expected to eventually power most Google services, as Gemini 3 did.

One independent review found no Argon model ID in the public Gemini API catalog, so don’t assume a subscription gives you access today. Google gave no date for the wider release.

Official Statements

Tulsee Doshi, Google’s Gemini model product lead, called Argon “incredibly well-rounded” and strong at long, multi-step tasks. She said the staged rollout puts a cyber-defense model in defenders’ hands early.

Google also says it is working with the US government’s voluntary pre-release testing program for frontier models.

Competitive Pressure 

Argon puts real pressure on OpenAI and Anthropic. Vals AI said Gemini topped its Vals Index for the first time.

Some caution is fair. Before launch, Bloomberg reported that some people inside Google worried Argon wasn’t as strong as rival models. Real-world testing will matter more than launch charts, and that can’t happen until access widens.

Final Thoughts

Argon looks strong on paper, especially for business and security work. For now, though, it’s a model most people can read about but not try. Watch for the Ultra and API rollout date and for independent test results.