Google has announced Gemini 4 Argon, the model it describes as its next era of frontier intelligence, and almost nobody can use it yet. In its announcement, Google says Argon is rolling out first to a group of trusted cyber defenders through its Fairwind Program, with developers, enterprises and consumers to follow as soon as possible.
The price is already set, and it puts Google squarely in this week’s fight. At an introductory $2.81 per million input tokens and $14.04 per million output tokens, Argon costs the same as Anthropic’s Claude Sonnet 5.5 and OpenAI’s GPT 6.1 Sol, both released this week. After that period it rises to $5.62 and $28.08, the price of Claude Opus 5.5. Google bills the API in US dollars, so the local figure moves with the exchange rate.
Who can use Argon today
For now, Argon is limited to the Fairwind cohort and Google’s own teams. Google says it is taking a phased approach, is working with the US government’s voluntary process for pre release model access, and will open the model first to paid API customers and Google AI Ultra subscribers. It has not given a date. In Australia, Google AI Ultra starts at $149.99 a month.
Trusted defenders receive Argon without its cyber guardrails, so they can use its full ability to find, validate and patch software vulnerabilities. Everyone else will get a version designed to refuse harmful requests, with monitoring of the model’s reasoning and actions that can stop a task when it oversteps what the user asked for.
What the independent numbers say
Artificial Analysis has already scored Argon on its Intelligence Index, testing it at high effort. It scores 52.6, ahead of GPT 6.1 Sol at its maximum setting (51.8) and level with OpenAI’s flagship GPT 6 Astra (52.7), but behind Claude Sonnet 5.5 at maximum effort (56) and Claude Opus 5.5 (57.6).
Cost per task is where the three models at the same token price split apart. Running the index cost about $2.79 per task on Argon, against $1.02 for GPT 6.1 Sol and $10.70 for Claude Sonnet 5.5 at their maximum settings, because each model burns a different number of tokens to reach an answer. Argon’s speed has not been measured yet.

What Google says Argon does well
Google’s own figures are strongest on long, multi step work. It reports 77.9% on DeepSWE v1.1, a test of long software engineering tasks, first place on the Vals Index of finance, legal, tax and coding work, and 51.3% on Zapier’s AutomationBench. It has also raised the output limit to one million tokens per response, up from 64,000.
Inside Google, Argon agents are migrating C and C++ code to Rust, including more than 800,000 lines of the Fuchsia Zircon kernel, and have freed more than 300 TiB of memory across Google’s data centres. These are vendor claims rather than independent tests, and they come from a company running the model on its own systems.
Why Australian businesses should care
The practical news is the price. For now, three frontier models sit at the same token rate, so the choice comes down to quality, speed and how many tokens each one uses, not the rate card. For teams already on Gemini through Workspace or Google Cloud, Argon is the upgrade to plan for, but not one to build on until Google opens access.
Until then, the useful step is to compare what you can run today. Our Claude vs Gemini comparison covers the models Google sells now, including Gemini 3.1 Pro, which Google still calls its leading model in the Gemini app.
The security angle matters locally too. Argon’s headline skill is finding and patching vulnerabilities, the same kind of capability behind last week’s admission that OpenAI’s models reached four Australian government sites during training. Google giving defenders the model first is a deliberate choice, and a reminder that agent access to company systems needs limits.
A slower release by design
Google has not said how long the Fairwind phase will last or when Australian customers will see Argon in the API. When it arrives, the first question for local teams will be cost per task rather than price per token. That is the number to test before moving any workload, and it belongs in any AI strategy review this quarter.
