Gemini 4 Argon: what Google's new model is and when you can use it

What happened

On September 30, Google announced Gemini 4 Argon, the first model of the Gemini 4 generation. Argon is its official name, not a leaked codename, and it's the only Gemini 4 model Google has named. There is no Gemini 4 Pro or Flash yet.

It isn't in the Gemini app or the public API. Google is rolling it out in phases, starting with trusted cyber defenders in its Fairwind Program, plus Google's own teams. Paid API customers and Google AI Ultra subscribers come next. Google says only that this will happen "as soon as possible", with no date. As of October 13, the Gemini API's model list, pricing page and changelog have no Argon entry, and neither do the Gemini subscription plans.

What Google says it's good at

• Coding. 77.9% on DeepSWE v1.1, ahead of Claude Opus 5.5 (74.2%) and GPT-6 Astra (74.1%) in Google's table.
• Legal and finance work. First place on AutomationBench (51.3%) and the Vals Index, and strong results on multi-step financial and legal research.
• Long video and charts. 91.7% on LVBench, a long-video understanding test.
• Cyber defense. Finding, checking and patching vulnerabilities, which is why defenders got it first. It ties for first on CWE-bench v1 at 68%.
• Long answers. Output goes up to 1 million tokens, from 64,000. Google doesn't state the input context window.

These are Google's own numbers, and the same table shows where Argon loses: GPT-6 Astra leads on FrontierSWE v2 and OSWorld-2.0, and Opus 5.5 leads on Terminal-bench 4.0. On the independent Arena text leaderboard, a high-effort Argon sits at number one, ahead of Opus 5.5, with the score still marked preliminary.

What it will cost

Google has already published the API price: $2 per million input tokens and $10 per million output tokens for an introductory period, then $4 and $20. Cached input is 95% off. Google hasn't said how long the introductory price lasts. For comparison, Gemini 3.1 Pro Preview costs $2 input and $12 output today for prompts up to 200,000 tokens.

On the consumer side, Google AI Ultra costs $99.99 or $199.99 a month in the US, so the cheapest announced route to Argon in an app is not cheap.

Rumors you'll see

• "Carbon". Business Insider reported employees testing a stronger internal Gemini 4 checkpoint called Carbon. Google hasn't commented, and it's unclear whether it would ship as an Argon update or a separate model.
• Model IDs. Some guides print IDs such as "gemini-4-argon". Google hasn't published any, so code that uses them won't work.
• A Meta model that beats it. Posts about Meta's "Watermelon" model claim a higher score. As of October 13, Meta hasn't released it or published benchmarks.

Which Gemini to use now

• Free plan: Gemini 3.5 Flash-Lite.
• Google AI Plus ($4.99): 3.6 Flash and 3.1 Pro.
• Google AI Pro ($19.99) and Ultra: 3.8 Flash and 3.1 Pro, with a 1 million-token context window on Pro.
• API: gemini-3.8-flash at $0.75 input and $3.75 output until December 31, with a free tier, or gemini-3.1-pro-preview for harder tasks.

Prompts that work well on 3.1 Pro should carry over: long, specific, with the files attached. Our Gemini prompts are written that way, and the Gemini review compares the plans.

Try it

Two prompts that test what Google says Argon is best at. They run on Gemini 3.1 Pro today, so you'll have a baseline to compare once Argon reaches you.

I've attached a [length, e.g. 90-minute] video of [what it is, e.g. a product team's planning meeting]. Watch all of it. Then give me: 1) a timeline of the main sections with timestamps, 2) every decision that was made, with the timestamp and who made it, 3) every open question nobody answered, 4) anything shown on screen (slides, charts, numbers) that matters and isn't said out loud. Quote only what is actually said or shown. If audio or video is unclear somewhere, give the timestamp and say so instead of guessing.

Act as a senior application security engineer. Review the code I've attached, written in [language and framework], which [what the code does, e.g. handles login and password resets]. List every vulnerability you find, most serious first. For each one: the file and line, what an attacker could do, how confident you are (high, medium or low), and a fixed version of the code. Then list the checks you couldn't do without more context, such as config files or how the code is deployed. Don't report style issues.

Sources

• Google, Gemini 4 Argon: our next era of frontier intelligence, September 30, 2026: name, access, API price, output limit, strengths
• Google DeepMind, Gemini models: benchmark table
• Google, Gemini API pricing and changelog, updated October 9, 2026: current models and prices, no Argon entry
• Google, Gemini subscriptions, checked October 13, 2026: models on each plan
• 9to5Google, Google says Gemini 4 release is coming as soon as possible, September 24, 2026
• Arena, Text leaderboard, updated October 8, 2026
• The Decoder, Meta's Watermelon model due in October