Gemini 4 Argon: Google's Model That Finds and Fixes Flaws

Short answer: Gemini 4 Argon is the first model of Google's Gemini 4 generation, announced on September 30, 2026, for long and complex work in software engineering, legal and finance research, and cybersecurity defense. Its headline skill is security: Google says Argon can find, validate and patch critical software vulnerabilities on its own. Because the same skills could help attackers, Google released it first to vetted cyber defenders through its Fairwind Program (Google). On October 13, 2026, Argon became globally available to developers and enterprises through the Gemini API, to Google AI Ultra subscribers in Google Antigravity, and to Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search (Google, October 13, 2026). It's no longer Google's newest model, though. Gemini 4 Carbon, which Google calls its strongest model, has been available since November 10, 2026, to paying developers in the Gemini API, to businesses in Gemini for business, and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity (Google, November 10, 2026). Google still recommends Argon for most coding and knowledge work, and Carbon for the hardest multistep tasks.

Most frontier models open to everyone on launch day. Argon didn't. Google treated its first Gemini 4 model as a tool that can harden software or break it, depending on who holds it, and it decided who went first. This guide covers what Argon does, how its vulnerability work is described, why access started restricted, who can use it now and what it costs. It also explains what Carbon adds, who can use it now and when it's worth picking over Argon. For a wider view of assistants and agents, start with our AI tools hub.

What is Gemini 4 Argon?

Gemini 4 Argon is the first model of Google's Gemini 4 generation. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect, introduced it as a model built for deep reasoning across long workflows: tasks that take many steps and a lot of context, such as migrating a large codebase or producing a full legal or financial analysis.

The most visible technical change is output length. Argon can write up to 1 million tokens in a single response, up from the 64,000 token limit of earlier Gemini models. In practice, that lets the model deliver a large piece of work in one pass instead of splitting it into dozens of smaller requests that you have to stitch back together.

Google backs the claim with internal examples. According to the announcement, Argon helped migrate more than 800,000 lines of code for the Zircon kernel in Fuchsia and made a video decoder 2.7 times faster.

What can Argon do with code and long tasks?

Google published scores across coding, automation and enterprise work. These are Google's own reported results, not independent tests:

On CWE-bench v1, Argon shares first place with OpenAI's GPT-6 Astra and xAI's Grok 4.7, according to SecurityWeek. Google also says Argon leads Gray Swan's indirect prompt injection benchmark, which tests how well a model resists instructions hidden in the content it reads. That matters for any agent that browses the web or reads tickets and email.

On the enterprise side, Google reports leading results on the Vals Index, the Vals Finance Agent benchmark and Harvey's Legal Agent Benchmark.

How does Argon find and patch vulnerabilities?

Google describes Argon's security work in three verbs: find, validate and patch.

• Find: read a codebase and spot code that could be exploited, from a missing input check to a flaw in how memory is handled.
• Validate: confirm that the flaw is real and reachable, so the team isn't buried in false alarms. A finding you can reproduce is worth far more than a long list of maybes.
• Patch: propose a code change that closes the hole without breaking what the code is supposed to do.

Google tested this on an internal benchmark of complex codebases written in 20 programming languages. The most concrete example came from Wiz, a cloud security company, through its Scan for Good initiative: Argon found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals around the world, a flaw that previous frontier models had missed. The vendor and the patch status were not disclosed (Help Net Security). On Wiz's black box penetration testing benchmark, Argon also beat Gemini 3.8 Flash Cyber at mapping what an attacker could reach and at producing proofs of concept.

Argon builds on a model Google released a few weeks earlier. In September, Gemini 3.8 Flash Cyber went to vetted defenders, and Google's Chrome Security team reported that it produced correct vulnerability patches at 2.6 times the rate of the best comparable commercial models (Tech Times).

Why did Google restrict access to Argon?

Because security skills cut both ways. A model that can find a flaw, prove it works and write the fix can also hand an attacker a working exploit. Google chose to give defenders a head start:

• Fairwind Program first. Google launched Fairwind on September 3, 2026, alongside Gemini 3.8 Flash Cyber, for government agencies, critical infrastructure operators, maintainers of widely used software and vetted security partners. Argon went to these trusted defenders first, and on October 13, 2026, Google began expanding it to hundreds of additional Fairwind Program partners (Google).
• A version without cyber guardrails. Fairwind members and Google's own internal teams get Argon without its cyber restrictions, so they can use its full defensive capabilities.
• Government review. Google says it is taking part in the U.S. government's voluntary process for giving it model access before release.

The public version comes with safeguards. Google says Argon is designed to refuse requests that would enable cyber, chemical, biological, radiological or nuclear attacks while still supporting legitimate research that could serve both sides. Google's announcement also lists monitoring of the model's internal activations, automated red teaming against prompt injection, monitoring of its chain of thought for misaligned behavior and hardened sandboxes for agents.

OpenAI took a different route with GPT-6 Astra, released on September 3. Astra reached ChatGPT Plus, Pro, Business and Enterprise users and the API within days, but it refuses more advanced cybersecurity tasks, such as writing proof of concept exploits. OpenAI says it plans to widen access with less restrictive safeguards for defensive work through its Daybreak program (OpenAI). Put simply, OpenAI released the model and held back the risky capabilities, while Google held back the model itself until October 13, 2026.

Who can use Gemini 4 Argon today?

Google confirmed general availability in its post "Gemini 4 Argon is now available for everyone" (Google, October 13, 2026). Developers can call Argon through the Gemini API today. Consumers need a Google AI Ultra or Pro subscription to use it in the Gemini app and AI Mode, and Google AI Ultra to use it in Antigravity.

How much will Gemini 4 Argon cost?

For developers, Google has published API prices. Argon launched on October 13, 2026 at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper than regular input. After the introductory period ends on November 30, 2026, the price rises to $4 per million input tokens and $20 per million output tokens (Google, October 13, 2026).

For consumers, Argon is included with Google AI Ultra and Pro subscriptions in the Gemini app and AI Mode in Search, and with Google AI Ultra in Google Antigravity.

Keep the 1 million token output limit in mind when you estimate costs. A single long answer can consume a lot of output tokens, so test on real tasks before you plan a budget.

What is Gemini 4 Carbon, and should you use it instead of Argon?

Gemini 4 Carbon is the model Google announced on October 28, 2026, two weeks after Argon, and opened up on November 10, 2026. Think of it as the next step up in the same family. Google calls it its strongest model, aimed at the same two jobs as Argon: defending software and handling long engineering tasks (Google, October 28, 2026).

Here's what Google says Carbon adds. All of these are Google's own reported results, not independent tests:

Google also says Carbon generates secure code at more than twice Argon's success rate on its internal tests. It says Carbon makes about half as many unsupported claims as Argon. And its teams have watched Carbon keep working for days on hard, multistep engineering problems, like tracking down subtle bugs across large codebases.

Who can use Carbon today? Since November 10, 2026, a lot more people, but not everyone. Here's where Google says it's available (Google, November 10, 2026):

Notice what's missing. Google AI Pro isn't on the list, and the API route needs the paid tier. If you're on Pro or use the API without paying, you're still on Argon for now.

What does it cost? In the API, Carbon has the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. When that period ends on November 30, 2026, it moves to $4 per million input tokens and $20 per million output tokens. Google said in its October 28 announcement that cached input is 95% cheaper than regular input. One catch: Google says Carbon uses more tokens than Argon because it explores more and runs longer checks on its own work. The same price per token can still mean a bigger bill per task.

What's /swarm? Alongside Carbon, Google started previewing /swarm in Google Antigravity. You give one prompt, and a team of agents splits the work, runs it in parallel and checks each other's changes, with Carbon in the roles that need the deepest reasoning. Antigravity's /swarm docs say it runs in the Antigravity App and CLI on Google AI Pro and Ultra plans, and that a swarm uses your quota much faster than a single conversation. Carbon itself is only for Ultra subscribers in Antigravity.

So should you wait? There's nothing to wait for anymore. Google's own advice is that Argon stays the right choice for most software engineering and knowledge work, and Carbon is for tasks that push a model to its limit, like complex multistep debugging or open ended research. If you're testing on real code today, start with Argon. Then run your hardest cases through Carbon and compare the results and the token count before you switch.

What does Argon mean for developers and security teams?

Models like Argon and Astra change how software gets secured. A few practical consequences:

• More bug reports found by AI. Maintainers of popular projects should expect more vulnerability reports produced with AI. Validated findings help, but triage still takes people and time.
• AI patches still need review. A patch that closes one hole can open another or quietly change behavior. Run tests, check the diff and keep a human in the approval loop. Our guide on how to review and verify AI generated code covers the checks that matter.
• Basics matter more, not less. When models find the easy flaws in minutes, clean input validation, current dependencies and least privilege stop being optional.
• Agents need limits. A model with access to your code and tools should run in a sandbox, with scoped permissions and logs. The OpenBot governance case shows what that isolation looks like in practice.

How can you start using Argon?

1. Check whether Fairwind applies to you. If your organization is a government agency, a critical infrastructure operator or the maintainer of widely used software, Fairwind is the route to the version of Argon without cyber guardrails.
2. Build a small test set. Collect a few real bugs and vulnerabilities your team has already fixed. Run them through Argon in the Gemini API to see whether it finds them and how good its patches are.
3. Set up review gates now. Automated tests, static analysis and a human approval step should sit between any AI patch and your main branch.
4. Compare it with other tools. Other coding agents already handle a lot of this work. Our comparison of the best AI coding agents shows which tools fit which jobs.

The skill behind using a model like Argon

Argon doesn't remove the need for people who understand code and security. It rewards them. Someone still has to decide what to scan, judge whether a finding is real and approve a patch before it ships. Those are skills you can learn. If you want to build them with mentors and real projects, compare our AI programs and find the one that fits your starting point.