Google has officially unveiled Gemini 4 Argon, its latest Frontier AI model, promising major improvements in coding, long-term tasks, professional knowledge work and cyber security. But unlike a typical Gemini or AI model launch, you probably won’t be able to try this right away.
Announced by Google earlier today, Gemini 4 Argon will initially be made available to a select group of trusted cybersecurity defenders through Google’s Fairwind program. Greater availability for developers, businesses and consumers will come later, starting with paid API customers and Google AI Ultra subscribers. Google hasn’t given us an exact date yet.
And I can understand why the company is taking a more cautious approach this time.
Gemini 4 Argon can continue to work for much longer
One of Argon’s most noticeable improvements is its huge 1 million token output limit, up from 64,000 tokens previously. Google says the extra seating space allows Argon to reason through extremely long and complicated tasks without running out of space halfway through them. Its input context window is also one million tokens, according to early technical analysis.
This feels particularly relevant as AI increasingly moves from responding to individual requests to doing longer jobs autonomously. Google says thousands of its own employees already use Argon internally. Some of the examples are significantly more ambitious than asking Gemini to rewrite an email.
Argon agents were involved in migrating C and C++ code to Rust, including working with more than 800,000 lines of code in the Fuchsia Zircon kernel. Another group of agents analyzed Google’s data center telemetry and identified memory optimizations that reportedly freed up more than 300 TiB of memory after deployment. Google also claims Argon has helped improve a quantum computer algorithm over a previously published baseline.
Of course, these are Google’s own examples, so I’m still interested to see how Argon behaves when developers outside of Google start throwing messy real-world projects at it.
Google has specially trained Gemini 4 Argon for defensive cybersecurity work. The company says the model can autonomously find, validate and patch critical software vulnerabilities.
On CWE-Bench v1, the Argon scored 68%, the highest score according to Google’s results. Security firm Wiz also tested Argon through its Scan for Good initiative, with Google saying the model discovered a critical vulnerability affecting health software that previous frontier models missed.
But these capabilities are also part of the reason why Google isn’t simply switching Argon inside Gemini for everyone today. Google says it strengthens security against malicious cyber use, CBRN-related abuse, indirect prompt injection attacks and models that take actions beyond a user’s intentions. There are also test systems that monitor Argon’s reasoning and actions and can stop execution if necessary.
We are reaching a point where the headline feature of a new AI model is not simply that it can write or reason better, but that companies are increasingly thinking about what happens when these models can independently perform complex technical tasks for long periods of time.
Google’s benchmarks naturally paint a very impressive picture. Argon scores 77.9% on DeepSWE v1.1 for long-horizon software engineering, tops the Vals Index covering professional work such as finance, coding, legal and tax tasks, and scores 91.7% on LVBench for long-video understanding.
Independent tests give us a bit more perspective. Artificial analysis reportedly ranks Gemini 4 Argon at 53 on its intelligence index, putting it roughly alongside other current frontier models rather than clearly ahead of everything else. Interestingly, Argon seems to use significantly more output tokens to complete some tasks, so the huge output allowance can make quite a lot of work.
Google will initially price Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens, before eventually raising these rates to $4 and $20 respectively.
Gemini 4 Argon looks like an important step towards AI models that can not only answer tougher questions, but can focus much longer on complicated jobs. The bigger question is when Google will feel comfortable enough to let the rest of us use it.