Google Just Released Gemini 4 Argon: How to Make the Most of It

After months of delays, Google has shipped its new flagship model, Gemini 4 Argon. It was released to a small group of trusted cybersecurity organisations on 30 September 2026 and opened to everyone on 13 October.

Here's what changed, where Argon is genuinely strong, and how to get the most out of it without paying for its weak spots.

Where you can use it

• Developers: the Gemini API and AI Studio, with the model ID gemini-4-argon.
• Google AI Pro and Ultra: the Gemini app and AI Mode in Search.
• Google AI Ultra: also Antigravity, Google's agentic coding environment.

What it costs

	Input (per 1M tokens)	Output (per 1M tokens)
Introductory, until 30 Nov 2026	$2	$10
Standard, from 1 Dec 2026	$4	$20

Cached input is 95% cheaper. For comparison, Claude Opus 5.5 lists at $4 and $20, and GPT-6 Astra at $10 and $50. If you want to test Argon properly, the introductory window is the cheapest time to do it.

The big change: a million tokens out, not just in

Older Gemini models could read a huge amount but only write about 64,000 tokens in one answer. Argon has a unified limit of about 1 million tokens shared between input and output.

In practice that means Argon can finish long jobs in one pass: rewriting a large codebase module, producing a complete report from hundreds of documents, or running a long multi-step agent task without stopping halfway.

Where Argon is strong

By Google's own published numbers (independent results are still limited):

• Long documents. On GraphWalks, a test of reasoning across 256,000 to 1 million tokens, Argon scores 84.2% against 71.8% for GPT-6 Astra and 66.8% for Claude Opus 5.5.
• Broad knowledge work. On the Vals Index it scores 68.9%, ahead of Opus 5.5 (67.0%) and Astra (63.1%).
• Software engineering on real repositories. It leads DeepSWE v1.1 at 77.9%.
• Video. 91.7% on LVBench, a long-video understanding test.

Google also shared internal results: in Chromium's 40-million-line codebase, Argon helped split out a roughly 6-million-line legacy folder, and agentic code submissions rose by more than 35%.

Where it's weaker

The same tables show gaps, mostly in terminal-driven work:

• Terminal-Bench 4.0: 57.4% versus 66.4% for Opus 5.5.
• FrontierSWE v2: 55.0% versus 65.5% for Astra.

Industry analyst Pareekh Jain summed it up for InfoWorld: strong at multi-step tool use and long documents, roughly average at everyday coding, and behind on creative writing and command-line work.

How to make the most of it

1. Give it the long jobs. Contracts, financial filings, research archives, entire code modules. Argon's advantage grows with the size of the input. Pasting a single paragraph wastes its strengths.

2. Ask for complete outputs in one go. "Produce the full migration plan with every file listed" now fits in one answer. You no longer need to ask it to "continue".

3. Cache what you reuse. If you query the same document set many times, cached input costs 95% less. Load the archive once and ask many questions against it.

4. Keep your current tool for terminal coding. If your team lives in the command line, keep your existing coding assistant for that and use Argon for long-context analysis and planning.

5. Use the introductory window to benchmark. Run 20 to 50 of your own real tasks through Argon and your current model before 30 November. Compare quality, speed and cost on your work, not the leaderboard.

6. Expect occasional slowness. Google's model card notes hallucinations and occasional slowness or timeouts. Build retries into anything automated, and check facts that matter.

Think of it like picking a Test cricket opener versus a T20 finisher. Argon is built for the long innings. Use it where patience and stamina win, and don't judge it on the format it wasn't picked for.

In one line: Gemini 4 Argon shines on long documents and long tasks thanks to its 1M input-output limit, so give it the big jobs, cache your inputs, and keep your current tool for command-line coding.