What you need to know
- Gemini 4 Argon is Google’s new frontier model for long, multi-step work in software engineering, enterprise knowledge work and defensive cybersecurity. Google announced it on September 30, 2026. The important catch is availability: Argon is not a normal public Gemini model yet. Its first rollout is to trusted cyber defenders through Google’s Fairwind Program while broader safeguards are tested.
- The headline change is not one benchmark. It is the attempt to keep a model working coherently across much longer tasks. Google says Argon raises the output-token limit from 64,000 to 1 million tokens. That is an output limit, not merely a context-window marketing number, and it is aimed at jobs that may require sustained reasoning, code changes, research and repeated tool use.
- For most people, no. The first external users are a limited group of cyber defenders in the Fairwind Program. Google says the phased release is deliberate because stronger autonomous cyber capabilities create a harder safety problem than a conventional chatbot launch.
The short answer
Gemini 4 Argon is Google’s new frontier model for long, multi-step work in software engineering, enterprise knowledge work and defensive cybersecurity. Google announced it on September 30, 2026. The important catch is availability: Argon is not a normal public Gemini model yet. Its first rollout is to trusted cyber defenders through Google’s Fairwind Program while broader safeguards are tested.
Google says developers, enterprises and consumers will get access later, beginning with paid API customers and Google AI Ultra subscribers. That makes the present moment unusual: the model is announced, benchmarked and already being used inside Google, but general users cannot simply select it in Gemini today.
What actually changed in Gemini 4 Argon?
The headline change is not one benchmark. It is the attempt to keep a model working coherently across much longer tasks. Google says Argon raises the output-token limit from 64,000 to 1 million tokens. That is an output limit, not merely a context-window marketing number, and it is aimed at jobs that may require sustained reasoning, code changes, research and repeated tool use.
Google describes internal uses that make the intended direction clearer. Argon has been used on large C and C++ to Rust migrations, specialized coding work, research and data-centre optimisation. Google says one internal agent effort identified memory savings already freeing more than 300 TiB, with a larger potential saving if fully rolled out.
For software engineering, Google reports 77.9% on DeepSWE v1.1. For vulnerability remediation, it reports 68% on CWE-bench v1. Those numbers are vendor-reported results and should be treated as evidence to investigate rather than a guarantee of performance in a particular company’s codebase.
Can you use Gemini 4 Argon right now?
For most people, no. The first external users are a limited group of cyber defenders in the Fairwind Program. Google says the phased release is deliberate because stronger autonomous cyber capabilities create a harder safety problem than a conventional chatbot launch.
There is also no reason to build production software around an assumed public model identifier before Google publishes one. Teams interested in Argon can prepare their evaluation harness, data boundaries and model abstraction now, then substitute the official model when access actually arrives.
Google says broader availability will start with paid API customers and Google AI Ultra subscribers. Until that happens, any website claiming that everyone can already call a public Argon API deserves careful verification.
Gemini 4 Argon pricing
Google announced an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input priced at a 95% discount to the input rate. After the introductory period, Google says the price will rise to $4 per million input tokens and $20 per million output tokens.
The unusually large possible output makes cost modelling important. A model capable of generating extremely long trajectories can also create extremely large bills if an application allows unnecessary reasoning or output. Real cost will depend on task design, caching, tool calls, retries and how often a long-running agent succeeds without human intervention.
Why cybersecurity is central to the launch
Google is positioning Argon as a defensive cybersecurity model capable of finding, validating and patching software vulnerabilities. It says Wiz has been testing Argon through its Scan for Good initiative and that the model found a critical issue in healthcare software that earlier frontier models had missed.
Those capabilities explain part of the restricted rollout. A system that can autonomously reason about vulnerabilities is useful to defenders and potentially useful to attackers. Google says it is strengthening misuse controls, prompt-injection resistance, monitoring and sandboxing before broad access.
For ordinary users, this means the slow rollout is not evidence that the announcement is incomplete. The rollout itself is part of the product story.
Should developers and businesses care now?
Yes, but preparation is more useful than hype. If your work involves large codebases, long research tasks, complex financial or legal analysis, or security workflows, Argon’s direction is relevant. The real test will be whether it can complete those jobs reliably enough to reduce total human effort, not whether it can produce an impressive first answer.
Prepare representative tasks and acceptance criteria now. When access opens, compare Argon against the model and human workflow you already use. Measure completion quality, time, cost, error recovery and supervision required. That will tell you more than a launch-day leaderboard.