Gemini 4 Argon Pricing: $2 Intro Rate, Then It Doubles

Gemini 4 Argon pricing has a catch printed right in the announcement: $2 per 1M input tokens and $10 per 1M output tokens now, and $4 and $20 later. Google hasn't said when "later" is. If you plan security tooling around the intro rate, you are planning around a number with no expiry date.

This piece covers what was announced, what you can't use yet, and how to budget so the price doubling doesn't surprise you.

What Google announced about Gemini 4 Argon

On September 30, 2026, Google announced Gemini 4 Argon as a phased rollout. Coverage of the announcement, including Tbreak and Technology.org, reports that the first users are trusted cyber defenders and testers in the Fairwind Program.

The numbers, per 1M tokens:

| | Introductory | Standard | | --- | --- | --- | | Input | $2 | $4 | | Output | $10 | $20 | | Cached input | 95% off input price | 95% off input price |

The maximum output length rises to 1M tokens, up from 64K, according to SmartScope's breakdown of the release. Long outputs are where your bill will come from, since output costs five times input.

Gemini 4 Argon access: who can use it today

Almost nobody. Per SmartScope, as of October 1, 2026 it is limited to Fairwind testers, and even Google AI Ultra subscribers can't use it yet. Paid API customers and Ultra subscribers are next, with no date given.

Still unconfirmed in the sources I read: the end of the introductory price, the general availability date, the input context limit and the official API model ID. Independent benchmarks are also missing, so treat performance claims as vendor claims.

Budget for Gemini 4 Argon cached input and the price jump

Cached input is the lever. Reuse the same long prefix, such as a detection ruleset, a runbook or a system prompt, and that part costs 5% of the normal input rate. At the intro rate that is $0.10 per 1M tokens.

The sources say 95% off in both periods. At the standard rate that would work out to $0.20, but that is my arithmetic, not a published price.

Here is a small script to model both rates before you have access:

PRICES = {
    "intro":    {"in": 2.0, "out": 10.0},
    "standard": {"in": 4.0, "out": 20.0},
}
CACHE_DISCOUNT = 0.95  # 95% off the input price, per the announcement

def monthly_cost(tier, cached_in_m, fresh_in_m, out_m):
    p = PRICES[tier]
    cached = cached_in_m * p["in"] * (1 - CACHE_DISCOUNT)
    return cached + fresh_in_m * p["in"] + out_m * p["out"]

# Millions of tokens per month: 30M cached, 10M fresh input, 4M output
for tier in PRICES:
    print(tier, round(monthly_cost(tier, 30, 10, 4), 2))

For that workload it prints 63.00 at the intro rate and 126.00 at the standard rate. Everything doubles, so your budget should be written against the standard column. Treat the intro rate as a discount you may lose any day.

Where it fits for cyber and security teams

The model is aimed at defenders, and reports say Fairwind testers get access without the usual cybersecurity guardrails. That is a reason to be careful, not excited: decide who can call it, log every request, and keep it away from untrusted input until you have tested how it behaves.

Good candidates are long, repeated-context jobs: triaging alerts against a fixed playbook, reviewing code against a fixed policy, or summarizing long incident timelines. These benefit from caching. One-off short prompts won't.

When waiting is the right call

If you're not in Fairwind, don't build around this yet. The context limit and model ID are unconfirmed, there are no independent evals, and the price is temporary. A model you can't call is a poor basis for a contract.

Gotchas

• The intro period has no stated end date. Budget at $4 and $20.
• Output is priced at 5x input, and the new 1M output ceiling means one runaway response can cost real money. Set a max output cap.
• Cached input only saves money if your prompts share a stable prefix. Put the changing part last.
• Pricing was reported from announcement coverage. Check Google's own pricing page once access opens.

Budget at the standard rate, cache your long prefixes, and when Argon reaches your account you will already know what it costs.