Gemini 4 Argon Is Here: 1M-Token Output and Long-Video Analysis

The third major model launch in a month: after OpenAI’s GPT-6 family and Anthropic’s Claude Opus 5.5, Google introduced the first member of the Gemini 4 family, Gemini 4 Argon, on September 30. This time, though, the interesting part isn’t only the model. It’s how it ships: Argon opens to security teams first, not to everyone.

So the real question for brands is this: how does a model you can’t use today affect your website, your video and your workflows tomorrow?

What happened?

According to Google’s official announcement, Argon is built for long, multi-step professional work: software engineering, enterprise knowledge work (legal, finance) and cybersecurity. The highlights:

• Built for long work: Google positions the model for “complex, long-horizon workflows”.
• 1M-token output: The output limit jumps from 64K to 1 million tokens, so the model can produce very long documents, reports or code in a single pass.
• Long-video understanding: Argon reportedly sets the top score on LVBench, a long-video comprehension benchmark, at 91.7%. Google also highlights chart analysis and taking action based on a series of documents.
• Software and security: 77.9% on DeepSWE v1.1, which measures real-world software engineering, and a tie for first on CWE-bench v1, which measures vulnerability remediation, at 68%. TechCrunch reports that the model can autonomously find, validate and patch critical software vulnerabilities.
• Pricing: An introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period, prices rise to $4 and $20.

Citing the independent benchmarking firm Artificial Analysis, Engadget adds two more data points: Argon matches GPT-6 Astra on its composite Intelligence Index at around 60% of the cost per task, and its 15% hallucination rate is the lowest among leading models.

Why can’t everyone use it?

For now, Argon is rolling out only to trusted cyber defenders through Google’s security initiative, the Fairwind Program. Next in line are paid API customers and Google AI Ultra subscribers, followed by developers, enterprises and consumers. 9to5Google notes that no date has been given.

The reasoning is easy to follow. A model that can find vulnerabilities on its own is as risky in the wrong hands as it is valuable for defense. That’s why Google stresses a phased release, internal monitoring for misuse and a voluntary pre-release review with the US government.

It’s a new signal for the industry. The most capable models are starting to arrive through controlled access first rather than for everyone on day one. All the more reason not to tie your plans to a single model’s launch date.

Three concrete areas for brands

Benchmark tables are interesting, but what matters to you is which kinds of work these capabilities unlock. We see three areas.

1. Long, large-scale transformation work

Google says it has used Argon internally on large migrations, such as moving a codebase of more than 800,000 lines to another language. The brand-side equivalent is familiar: moving an old site onto a new platform, restructuring hundreds of pages into a new CMS, producing multilingual versions. A 1M-token output limit makes it possible to run this kind of work as a whole rather than piece by piece.

But AI can only carry a structure that’s already orderly. On a site with inconsistent content fields and scattered page templates, even the strongest model returns a scattered result. That’s why, in our Webflow web design projects, we set up the CMS from day one to be readable and portable.

2. Long-video understanding

For video production, the standout number is 91.7% on LVBench. A model that understands long video can speed up work like finding a specific moment in hours of raw footage, tagging shots, drafting summaries and subtitles from interviews, or making old campaign footage in the archive reusable again.

The balance matters here: understanding video is not telling a story with it. Which shot carries the story, the rhythm of the edit, the colour and sound decisions are still the team’s job. The value of AI is freeing the team from searching the archive so they can spend that time on those decisions. That’s our approach in video production too: speed up the repetitive work, keep the creative calls with the team.

3. Security is now part of your website

Giving security teams first access also shows how fast vulnerability discovery is about to get. That speed will be in attackers’ hands as much as defenders’. Forms, integrations, third-party scripts and on-site chat assistants become surfaces that get scanned more often and more automatically.

Google also ranks Argon at the top for robustness against indirect prompt injection (steering a model with instructions hidden inside a page or document). If you’re adding an AI assistant to your site, which data it can reach is as much a design decision as which model you pick.

What it doesn’t change

• You don’t have access today. Argon is currently limited to selected security teams. For the coming months it makes more sense to build a setup that works with today’s models and can switch when needed than to wait for this one.
• Most of the numbers are Google’s own. Independent measurements such as Artificial Analysis send a positive signal, but a test with your own data beats any table.
• 15% hallucination is still 15%. Being the lowest in the industry is good news, but output that reaches customers, such as prices, contract terms or product details, should still go through human review.
• The price isn’t fixed. The introductory rate will double. Budget on the long-term price, not the launch offer.

What it means for your brand

Three major launches in a month point the same way: models are taking on longer work, with less supervision, at a lower unit cost. GPT-6 Sol and Luna lead on cost, Claude Opus 5.5 on efficiency, and Gemini 4 Argon on long-horizon work and security. OpenAI’s launch of Dots agents on GPT-6 Astra, a day before Argon, completes the picture: AI is moving from answering to doing the work.

The brands that come out ahead won’t be the first to use the newest model. They’ll be the ones whose foundations are ready:

• Organise your content: A consistent CMS structure, clean page templates and clear content fields are the ground any model can read and build on.
• Catalogue your video archive: Store raw footage with date, project and content details. When long-video models open up widely, an orderly archive turns straight into time saved.
• Make security a project line item: Review forms, integrations and third-party scripts on your site regularly. As attacks automate, “we’ll look at it later” stops being an option.
• Don’t lock into one provider: GPT-6 today, Gemini 4 tomorrow, someone else the day after. Build your architecture so you can switch models.

Frequently asked questions

When will we be able to use Gemini 4 Argon?

Google hasn’t given a date. The model is opening first to security teams in the Fairwind Program; next come paid API customers and Google AI Ultra subscribers, followed by developers, enterprises and consumers.

How much will Gemini 4 Argon cost?

The introductory rate is $2 per million input tokens and $10 per million output tokens, with 95% off cached input. After the introductory period, it rises to $4 per million input tokens and $20 per million output tokens.

Can Argon make videos?

The announcement is about video understanding, not video generation. Argon scored the top 91.7% on LVBench for spotting details in long videos. That helps with archive search, tagging and summaries. Making a film is a separate process and still the production team’s job.

Let’s map out where the new models can add value to your website, your video content and your workflows. From strategy to design, end to end, under one roof: get in touch.