AI Beat

AnalysisTechPolicy

Google's Gemini 4 Argon: a frontier model at a mid-range price, but not for you yet

Google says its new top model leads on coding, legal and finance work and cybersecurity, and it will cost no more per token than mid-range rivals at launch. First it goes only to vetted cyber defenders, as AI labs get more careful about who gets their most capable models.

Server rack with blinking green lights in a data centre
Photo: Domaintechnik on Unsplash

Key points

  • Google announced Gemini 4 Argon on 30 September. For now it is available only to vetted cybersecurity defenders through a programme called Fairwind; developers, companies and consumers will follow, starting with paid API customers and Google AI Ultra subscribers. Google gave no date.
  • The launch price is $2 per million input tokens and $10 per million output tokens, the same as OpenAI's mid-range GPT-6.1 Sol. Cached input costs 95% less. After an introductory period the price doubles to $4 and $20, level with Anthropic's Claude Opus 5.5.
  • Google says Argon leads on DeepSWE (long software-engineering tasks, 77.9%), on the Vals Index of finance, legal, tax and coding work, and on Zapier's AutomationBench (51.3%). These are Google's own figures; independent tests will follow.
  • It can write up to a million tokens in one answer, up from 64,000, so it can work on one problem for much longer.
  • Google is releasing a version without cyber guardrails to trusted defenders. Wiz says Argon found a critical flaw in hospital software used worldwide that earlier models missed.
  • The cautious rollout follows OpenAI's decision to shelve GPT-6.1 Astra over safety. For businesses the takeaway is practical: plan for frontier-level AI at mid-range prices, but don't build on Argon until it is generally available.

Google has a new top model, and the most striking thing about it may be the price. Gemini 4 Argon, announced on 30 September, is pitched as Google’s most capable model for coding, office work in law and finance, and cyber defence. At launch it will cost the same per token as the mid-range models of its rivals.

The catch: almost nobody can use it yet.

Who gets it, and when

For now Argon is going only to a group of vetted cybersecurity defenders through what Google calls the Fairwind programme, and to Google’s own staff. The company says it is taking part in the US government’s voluntary process for testing models before release, and will open Argon to developers, businesses and consumers “as soon as possible”, starting with paying API customers and subscribers to Google AI Ultra, its most expensive consumer plan. It gave no date.

That caution is now a pattern among the leading labs. Days earlier OpenAI shelved GPT-6.1 Astra after it failed the company’s own safety tests. Google says it is using the time before a broad release to harden safeguards against misuse for cyber and weapons attacks, against prompt injection (instructions hidden in documents or web pages that try to hijack the model), and against the model going beyond what users asked of it.

The price

Google will launch Argon at an introductory price of $2 per million input tokens and $10 per million output tokens. Cached input, text the model has already processed and can reuse, gets a 95% discount, which works out at $0.10 per million tokens. After the introductory period, which Google didn’t date, the price doubles to $4 and $20.

What a million output tokens cost: Gemini 4 Argon against rival models

Standard list prices per million output tokens, in US dollars. Argon is not yet generally available; Google announced its prices on 30 September 2026.

  1. GPT-6 Astra (OpenAI)$50
  2. Claude Fable 5.1 (Anthropic)$50
  3. Claude Opus 5.5 (Anthropic)$20
  4. Gemini 4 Argon, later$20
  5. Gemini 4 Argon, at launch$10
  6. GPT-6.1 Sol (OpenAI)$10

Source: Google; OpenAI and Anthropic pricing pages, checked 30 September 2026

At launch, that puts a model Google calls frontier-class at the price of OpenAI’s mid-range GPT-6.1 Sol, and at a fifth of what OpenAI and Anthropic charge for their most capable models. Even at the later price it matches Claude Opus 5.5. Low launch prices that rise later are becoming a Google habit: its Gemini 3.8 Flash also has an introductory price that doubles in January.

For a feel of what it means, try the numbers for your own workload in our AI API cost calculator.

What Google says it can do

Google’s claims, all from its own announcement:

  • Software engineering. 77.9% on DeepSWE v1.1, a test of long, real-world coding tasks, which Google calls a new state of the art. Inside Google, Argon agents are rewriting C and C++ code in the safer Rust language, including an 800,000-line operating-system kernel, with human review before anything ships.
  • Office work in law, finance and tax. First place on the Vals Index, which weights finance, coding, legal and tax tasks by their share of US GDP, and strong results on a legal benchmark from Harvey and on Zapier’s AutomationBench (51.3%), which tests end-to-end business tasks.
  • Long answers. Argon can produce up to a million tokens in one answer, against 64,000 before, so it can reason through a hard problem in one go rather than in pieces.
  • Video and documents. 91.7% on LVBench, a test of understanding long videos.
  • Cyber defence. It ties for first on CWE-bench, which tests fixing security flaws. The security firm Wiz says Argon found a critical vulnerability exposing personal data in hospital software used worldwide, which earlier models had missed.

Benchmarks chosen and reported by the company that made the model are a starting point, not a verdict. Independent tests will come once outside developers get access.

The cyber question

The most unusual part of the launch is that Google will give trusted defenders and its own teams a version of Argon without its cyber guardrails, so it can find and fix vulnerabilities without refusing. That is the logic of the moment: AI is getting good enough at attacking software that defenders need the same power first, as we described in how AI is being used against hospitals and banks. It also explains why the general release is slow: a model strong enough to defend critical systems is strong enough to attack them.

What it means for you

  • If you use AI at work through ChatGPT, Gemini or Claude: nothing changes yet. Argon will reach consumers through Google AI Ultra first.
  • If your company pays for AI by the token: frontier-level capability is getting much cheaper. Budget for that, but don’t build on Argon until it is generally available and tested independently, and remember the price doubles after the introductory period.
  • If you work in IT security: watch for Google’s Fairwind programme and Wiz’s results. Defensive AI is moving fast, and so are the attackers.

Comments

What do you think? No account needed. Please be kind; comments that attack people are removed.

We store only your name and comment. See our privacy policy.

    Sources

    1. Google: Gemini 4 Argon, our next era of frontier intelligence (30 September 2026)
    2. OpenAI: API pricing
    3. Anthropic: Pricing
    4. Google AI for Developers: Gemini API pricing