Gemini 4 Argon Is Now Available: Access, Pricing, and What Changed

On September 30, 2026, Google announced Gemini 4 Argon and limited it to trusted cybersecurity defenders in the Fairwind program. On October 13, 2026, Google made the public model available worldwide, with the Gemini platform model ID gemini-4-argon. This page covers what changed, who can use it, the price through November 30, and the quality claims Google has published.

This is an independent site, not an official Google page. Benchmark and internal figures below are Google’s.

What changed since September 30

Our October 2026 update said developers could not use Argon yet. That was correct on September 30. It is out of date now.

The October 13 posts are Introducing Gemini 4 Argon by Koray Kavukcuoglu, SVP, Google DeepMind, and Gemini 4 Argon on Google Cloud by Michael Gerstenhaber.

Argon is a frontier reasoning and text model. Google’s demos show it wiring car-engine audio from Gemini Omni Flash in Three.js, and a video that uses Gemini 3.8 TTS and Lyria 3.5. Those outputs come from the other models. For video, start with what Gemini Omni is. The Omni API developer guide is about video, not Argon.

The Cloud post puts Argon in a four-part lineup: Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, and Gemma for open-weights on the edge. It also says model intelligence underpins the Gemini agent from Gemini at Work, covered in what Google Cloud announced on October 8.

Google calls Argon “just the beginning” of the Gemini 4 family. No dates are given for the other Gemini 4 models. The Gemini 4 roadmap preview is background on pre-training, not a release calendar.

Where each audience can use it

The token prices below apply to the public-access model. The scientist tier is an application, not an automatic grant.

Model ID and the 1M token limit

On the Gemini platform API the ID is gemini-4-argon. That is also the model in Google AI Studio.

Google says the API gives Argon an industry-leading 1 million unified input-output token limit. The launch posts do not break down how that budget splits between input and output.

Docs describe a continuation protocol for that limit. The launch posts give no steps for it. Read Google’s developer docs before you design a job that has to cross the boundary.

Pricing through November 30, and after

Input and output both double when the intro period ends. The 95% cached-input discount is stated through November 30, 2026. The rates Google published for after that date are $4 and $20.

Benchmarks and internal results Google reports

Per Google, from the October 13 posts. This site has not re-run them.

Google also reports these internal results:

• Antigravity: agentic code submissions up by more than 35%, code rollbacks down by nearly a third.
• Chromium: a Chrome refactor in a codebase of about 40 million lines, described as months of work done in weeks.
• ADK (google/adk-python): in-memory session prep time cut 40% in long sessions; cold imports 21% faster.
• Gemma 4 E2B on an A100 via vLLM: 2.85× end-to-end inference speedup, from about 37,900 tokens/s to over 108,000 tokens/s. That figure is Gemma’s serving speed in Google’s report, not an Argon output rate.
• xTB: up to 2.3× faster structure prediction and 1.8× faster conformer search.

Safety and cyber access

Google says Argon was trained to respect system boundaries, report negative findings honestly, pause for human confirmation before irreversible steps, answer more directly with less hedging, and follow style instructions more closely.

Cyber access is split:

• Public Argon blocks offensive tasks such as pentesting and exploitation. Defensive use goes through CodeMender.
• Fairwind access is without those cyber guardrails. Google is adding hundreds of partners.

Offensive testing does not belong on the public model ID. Google says those tasks are blocked there.

Should you switch now?

Our advice, from the published facts. Not a Google migration order.

1. Move a text or reasoning call to gemini-4-argon if the only blocker was the September 30 Fairwind limit, and you are on the paid API tier or on Ultra or Pro for the consumer surfaces in the table.
2. Keep video, speech, and music on Omni, Gemini 3.8 TTS, and Lyria. Argon can sit beside them in a demo. It does not replace them.
3. Put November 30, 2026 on the calendar before a high-volume route. Intro rates are $2 and $10 per 1M tokens. After that, $4 and $20.
4. Keep Flash where you need speed and volume. Google’s lineup calls Argon the frontier reasoning model and Flash the speed and volume model.
5. Read the continuation protocol in Google’s developer docs before a 1M-token job. The blog posts name it and do not specify it.
6. Leave the rest of Gemini 4 off the schedule until Google publishes dates.
7. Re-test system prompts. Google says answers are more direct, hedging is lower, style instructions stick better, and the model should pause before irreversible steps.
8. Keep offensive cyber work inside Fairwind if you are in that program. Public Argon is the mitigated model.

Minimal Python call

Same google-genai shape as other Gemini platform snippets on this site. Auth stays whatever you already use for the Gemini API.

from google import genai

client = genai.Client()

response = client.models.generate_content(
    model="gemini-4-argon",
    contents="List the risks in this design, including the ones that fail.",
)

The prompt asks for negative findings on purpose. Google says the model is trained to report them honestly.

Related reading

• Gemini in October 2026 — the September 30 hold, when developers could not call Argon
• Gemini 4 roadmap preview
• Gemini Omni API developer guide
• What is Gemini Omni? — video generation, which is a different model