Google’s launch artwork for Gemini 4 Argon. Image: Google
Google has finally unveiled Gemini 4, and it is holding most people back from it. The new frontier model, called Gemini 4 Argon, is rolling out first to a small group of trusted cybersecurity defenders through what Google calls its Fairwind Program, with developers, businesses and consumers to follow “as soon as possible”.
The launch comes less than a week after Google DeepMind’s boss, Koray Kavukcuoglu, said Gemini 4 would arrive “much earlier” than the end of the year. He wrote Wednesday’s announcement himself.
Who can use Gemini 4 Argon? #
For now, almost nobody outside Google. Argon is going to vetted cyber defenders first, and Google says it will keep collecting feedback from early testers and adjusting its guardrails before a wider release. When that comes, it will start with paid API customers and subscribers to Google AI Ultra, its top consumer plan. Google gave no date. Google also says it is taking part in the US government’s voluntary process for testing models before release. That is the same kind of arrangement the White House asked of AI companies in Tuesday’s self-policing accord, which Google signed.
Prices are already set. Argon will launch at an introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period, that doubles to $4 and $20. Google hasn’t said how long the introductory period lasts.
A million tokens out in one go #
The headline technical change is how much Argon can write in one response. Google has raised the output limit to 1 million tokens, up from 64,000, so the model can think and write for far longer before it has to stop. That is aimed at long jobs such as rewriting large codebases or working through a research problem in a single run.
Google says thousands of its own staff already use Argon, and gave examples:
- Data centres: a team of Argon agents studied memory use across Google’s data centres and applied fixes that freed more than 300 TiB of memory, with 500 TiB to 1 PiB expected in total.
- Rust rewrites: Argon agents are converting C and C++ code to Rust across Google, up to more than 800,000 lines for the kernel of its Fuchsia operating system. Google says the rewrites are being audited and reviewed before they go live.
- Video decoding: on libgav1, Google’s open-source video decoder, agents replaced 32,000 lines of hand-tuned code in a Rust version and made it 2.7 times faster.
- Quantum computing: in one test it beat a published baseline for a quantum subroutine by 40% “in a matter of minutes”.
On benchmarks, Google says Argon sets a new best of 77.9% on DeepSWE v1.1, a test of long software engineering tasks, tops the Vals Index of finance, legal, tax and coding work, and ranks first on Zapier’s AutomationBench with 51.3%. These are Google’s own figures and haven’t been checked independently.
No cyber guardrails for the defenders #
Cybersecurity is why the defenders go first. Google says it trained Argon to find, confirm and patch serious software flaws on its own, and it will give trusted defenders and its own teams a version “without cyber guardrails”. Security firm Wiz has been using it to scan critical public infrastructure for free, and Google says the model found a critical flaw exposing personal information in healthcare software used by hospitals around the world, one that earlier models had missed.
That puts Google alongside Anthropic and OpenAI, whose most capable cyber models are also handed only to vetted security companies such as Palo Alto Networks. Google has its own reason to be careful: earlier this month it admitted that a Gemini agent had hacked three real companies during testing.
For everyone else, Google says Argon refuses help with cyber, chemical, biological, radiological and nuclear attacks, and that it watches the model’s internal activity to spot misuse. It also says it will monitor Argon’s reasoning and actions and stop it if it tries to go beyond what the user asked. Google says it was careful not to feed those findings back into training, so Argon doesn’t learn to hide its thinking, and urged the rest of the industry to keep models’ reasoning readable too.
Doubts inside Google #
Not everyone at Google is convinced. Some employees say Gemini 4 does well on benchmarks but struggles with some real-world coding tasks, Bloomberg reported on Wednesday, citing people familiar with the matter. Google disputes that description.
Why it matters #
Gemini 4 is Google’s answer to a month in which OpenAI and Anthropic set the pace, and its benchmark claims put it back at the front. But like OpenAI, which cancelled GPT-6.1 Astra over safety worries, Google is now releasing its best model slowly, and only to people it trusts.
Sources: Google announcement, Google DeepMind, Bloomberg (“Google Grapples With Employee Skepticism About New Gemini Model”, Sep 30).