Archive · 2 Oct 2026
Google Unveils Gemini 4 Argon With One-Million-Token Output
Google has introduced a new AI model called Gemini 4 Argon.
It can produce up to one million small pieces of text, called tokens, at once.
That gives it more space to work on huge computer programs, long documents and complicated jobs.
Google says its employees are already using it for coding, research, debugging and engineering.
The company also says Argon helped move computer code to Rust and made one video decoder run faster.
Argon can reportedly find and fix some serious computer security problems by itself.
This could help defenders protect software, but it could also help attackers find weaknesses.
Because of that risk, Google is first testing it with trusted cyber defenders and internal teams.
Google plans to offer it later to paid developers, businesses and Google AI Ultra subscribers, but has not given a final date for broad consumer access.
Gemini 4 Argon can generate up to one million tokens, compared with the previous 64,000-token limit.
Google says thousands of employees are using Argon for coding, debugging, research and large engineering projects.
An Argon-assisted Rust video decoder reportedly ran 2.7 times faster than an earlier Rust implementation while producing identical output.
Google says Argon can autonomously find, validate and patch critical software vulnerabilities, prompting a restricted initial rollout.
Access will expand from trusted cyber defenders to paid API customers and Google AI Ultra subscribers, with initial pricing of $2 per million input tokens and $10 per million output tokens.
- Who
- Google, its employees, trusted cyber defenders and cybersecurity company Wiz are involved in testing Gemini 4 Argon.
- What
- Google unveiled a frontier AI model that can generate up to one million tokens and handle extended coding, research, enterprise and cybersecurity tasks.
- When
- Google has announced the model and begun limited testing, but the articles provide no specific calendar date for wider access.
- Where
- Argon is being used internally at Google and tested through the company’s Fairwind Program; it is also being used by Wiz through its Scan for Good initiative.
- Why
- Google says the model is intended for long, complex workflows, while restricted access allows further testing of safeguards against cyber misuse, prompt injection and model misalignment.
Potential Benefits
Misuse Risks
Cybersecurity use
Potential Benefits
Google says Argon can autonomously find, validate and patch critical software vulnerabilities, helping trusted defenders identify threats and protect systems.
Misuse Risks
The same ability could help malicious users discover and exploit software weaknesses, creating a dual-use cybersecurity risk.
Restricted rollout
Potential Benefits
Google says phased access lets it test the model with trusted defenders and improve safeguards before wider availability.
Misuse Risks
Critics of broad access could argue that the model should remain limited until protections against cyber misuse, prompt injection and model misalignment are stronger.
Performance claims
Potential Benefits
Google reports strong results on software-engineering and automation benchmarks, as well as improvements in code migration and data-center memory efficiency.
Misuse Risks
The reported benchmark scores come from Google rather than independent assessors, so the results have not yet been independently verified in the articles.
- Model
- Gemini 4 Argon
- Maximum output
- One million tokens in a single output
- Previous limit
- 64,000 tokens
- Reported engineering result
- An Argon-assisted Rust video decoder ran 2.7 times faster than an earlier Rust implementation and produced identical video output.
- Reported benchmarks
- 77.9% on DeepSWE v1.1, 51.3% on Zapier’s AutomationBench and 68% on CWE-bench v1; these are Google-reported results, not independent assessments.
- Initial access
- Selected trusted cyber defenders through Google’s Fairwind Program, along with internal teams
- Pricing
- Initially $2 per million input tokens and $10 per million output tokens; later prices are planned to rise to $4 and $20, respectively, while cached input tokens receive a 95% discount.
