Gemini 4 Argon: Benchmarks, Pricing & Release Date
Gemini 4 Argon is Google’s new flagship AI model, and it has reopened the debate over who leads the AI race. Google announced it on September 30, 2026, as the top model of its Gemini 4 generation. However, the rollout is limited. Only Google’s own teams and vetted cybersecurity partners can use it today.
This guide explains what Gemini 4 Argon is, how its benchmark scores compare with GPT-6 Astra, what it costs, and when you might get access. Most figures come from Google’s own announcement, so treat them as company claims until independent testing confirms them.
What Is Gemini 4 Argon?
Gemini 4 Argon is the first model in Google’s Gemini 4 generation. It follows Gemini 3, which arrived last November. A Google spokesperson said Argon is larger than the company’s earlier advanced “Pro” models.
Google pitches the model for long, complex workflows. Its main targets are software engineering, legal and finance work, and cyber defense. Koray Kavukcuoglu, who now runs Google DeepMind day-to-day, called it the company’s “next era of frontier intelligence.” Sundar Pichai also announced the model on X.
The launch follows a rough stretch for Google. The company decided not to release the planned Gemini 3.5 Pro. It also changed leadership in its AI lab after months of delays. As a result, many observers see Gemini 4 Argon as a bid to catch up with OpenAI and Anthropic.
Why the Gemini 4 Argon Release Matters
Google’s position in AI has been questioned for months. Pichai pushed back on the idea that Google was losing ground in July. Still, rivals shipped new flagship models while Google delayed its own. Gemini 4 Argon is the company’s most direct answer so far.
The cybersecurity focus also stands out. Google says Argon found a software flaw that earlier models missed. In addition, Wiz is reportedly using Argon through the Fairwind Program to find critical vulnerabilities in hospital software. Those examples suggest Google wants Argon known as a defense tool first and a chatbot second.
Key Specs of Gemini 4 Argon
The biggest technical change is the output limit. Gemini 4 Argon can generate up to 1 million tokens in one response, compared with 64,000 for earlier Gemini models. The input context window stays at 1 million tokens.
This matters because long jobs no longer need to be split into many calls. For example, a developer could ask the model to handle a large code migration in one pass. Google says it already uses Argon internally to move code from C++ to Rust and to optimize storage in its data centers. According to Google, teams of Argon agents freed more than 300 TiB of memory across those data centers.
Thousands of Google employees use the model today. That gives Google real-world feedback. However, outsiders cannot verify the results yet.
Gemini 4 Argon Benchmarks vs GPT-6 Astra
Benchmarks are where Google makes its strongest case. By Google’s own figures, Gemini 4 Argon beats OpenAI’s GPT-6 Astra on 13 of 18 published benchmarks. Those include coding, automation tasks, and long video understanding.
Here are the headline results reported so far:
| Benchmark | Gemini 4 Argon | GPT-6 Astra |
|---|---|---|
| DeepSWE v1.1 (long coding tasks) | 77.9% | 74.1% |
| LVBench (video comprehension) | 91.7% | 87.5% |
| GraphWalks (long context) | 84.2% | 71.8% |
| Vals Finance | 65.4% | 53.5% |
| CWE-Bench v1 (security flaws) | 68% | 68% |
| Terminal-Bench Science | 57.6% | 68.1% |
| FrontierSWE v2 | 55.0% | 65.5% |
Google also reports that Argon scored 51.3% on AutomationBench, ahead of Claude Opus 5.5 at 42.5%. On DeepSWE, Opus 5.5 scored 74.2%, just above Astra.
Where Gemini 4 Argon Leads
Argon’s clearest wins come from long-context and professional tasks. It posted a large gap on GraphWalks, which tests long-context reasoning. It also led on finance and video benchmarks. In addition, its 77.9% on DeepSWE v1.1 sets a new high for that long software engineering test, according to Google.
Where GPT-6 Astra Still Leads
GPT-6 Astra stays ahead on Terminal-Bench Science and FrontierSWE v2. Argon also trailed on two of the four coding benchmarks in Google’s own release. On CWE-Bench, which measures security vulnerability detection, the two models tied at 68%.
So “matching” is a fair description. “Dominating” is not.
Gemini 4 Argon Pricing
Google is pricing Gemini 4 Argon aggressively. The introductory rate is $2 per million input tokens and $10 per million output tokens. Cached inputs cost $0.10 per million tokens. After the introductory period, prices rise to $4 and $20.
For comparison, OpenAI lists GPT-6 Astra at $10 per million input tokens and $50 per million output tokens. At launch, that makes Argon roughly five times cheaper. Even after the increase, it costs less than half of Astra’s rate.
Google has not said how long the introductory pricing will last. Therefore, budget for the higher rate if you plan a production project.
Gemini 4 Argon Release Date and Access
It is not publicly available yet. Google is rolling it out in stages:
- Today: Google’s internal teams and vetted cyber defenders in the Fairwind Program.
- Next: Paid API customers and Google AI Ultra subscribers.
- Later: Developers, enterprises, and consumers more broadly.
Google has not given dates for steps two and three. It says only that access will expand as soon as possible. According to one report, a senior executive used the phrase “well before the end of the year,” but that is not a firm commitment.
Fairwind is a program for vetted cybersecurity defenders, not a public waitlist. Members get Argon without its usual cyber guardrails so they can test attack detection and system hardening. Google says it is staging access because the model’s cyber abilities are strong. It is also joining the Trump administration’s voluntary pre-release model access process.
Developers should note that Google has not published an API model ID, so do not guess one. Google has not announced a free tier or free Gemini app access either.
Limitations and Early Skepticism
Treat the launch numbers with care. Google reported most of the benchmarks itself, and few people outside the Fairwind group have tested the model.
Bloomberg also reported that some Google employees doubt how well Argon handles real work. That does not disprove Google’s claims. Still, it shows why independent tests from groups such as Artificial Analysis will matter once access widens.
Finally, benchmarks measure narrow tasks. A model can lead a leaderboard and still feel slower or less reliable in daily use.
What Gemini 4 Argon Means for You
- Developers: Plan, but do not build on Argon yet. Set up a paid Gemini API project so you are ready when access opens. Until then, Gemini 3.8 Flash is the newest Google model you can put into production.
- Businesses: The pricing and long output limit could lower the cost of long-running agent workflows. Test Argon on your own tasks before you switch.
- Security Teams: Eligible organizations should look into the Fairwind Program, which is the only access route today.
- Everyday Users: Nothing changes yet.
Conclusion
Gemini 4 Argon shows that Google is back in the frontier race, at least on paper. It posts strong benchmark results, offers a 1 million token output limit, and costs far less than GPT-6 Astra at launch. Still, access is limited, the numbers are mostly self-reported, and there is no release date. Check back for updates as independent tests and the wider rollout arrive.
Gemini 4 Argon FAQ
Is Gemini 4 Argon available now?
Only to Google’s teams and vetted cyber defenders in the Fairwind Program. Paid API customers and Google AI Ultra subscribers come next, with no set date.
Is Gemini 4 Argon better than GPT-6 Astra?
On Google’s numbers, Argon wins 13 of 18 benchmarks. However, Astra leads on some coding and science tests, and independent results are still pending.
How much does Gemini 4 Argon cost?
The introductory price is $2 per million input tokens and $10 per million output tokens. It rises to $4 and $20 later.
Is Gemini 4 Argon free?
No free tier has been announced.







