Google announced Gemini 4 on September 30, 2026, in the form of its first model: Gemini 4 Argon. Google calls it a frontier model built for long, complex work in software engineering, legal and financial knowledge work, and cybersecurity defense.
The catch: almost nobody can use it yet. Argon is rolling out first to a select group of trusted cyber defenders through Google’s Fairwind Program, and Google says broader access will follow once it’s finished testing safeguards. No public release date has been announced.
Here’s what’s confirmed, what’s been reported, and what’s still unknown.
What Is Gemini 4?
Gemini 4 is the next generation of Google’s Gemini AI models, developed by Google DeepMind. Argon is the first Gemini 4 model Google has announced. As Axios put it, it’s Google’s long-awaited next-generation model, coming after a stretch in which the company mostly released smaller, cheaper Flash models while OpenAI and Anthropic pushed ahead.
A few names are easy to mix up:
- Gemini 4 is the model generation.
- Gemini 4 Argon is the specific model Google announced on September 30. Note the new naming style: no “Pro” or “Flash.”
- Gemini 3.x models (such as 3.1 Pro and 3.8 Flash) are the previous generation, still in use.
- Gemma 4 is a different product line: Gemma is Google’s family of open models, separate from Gemini. Other separately branded Gemini products aren’t the same thing as Argon either.
What Is Gemini 4 Argon?
Argon is a reasoning model, meaning it works through a problem step by step before producing an answer. Google says it was built to sustain deep reasoning across complex, long-horizon workflows, the kind of multi-step tasks that might take a person hours or days.
The announcement, written by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect, names the target areas: real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. In its closing section, Google adds creative writing to that list.
Why access is limited
The short answer is cybersecurity. Google says it trained Argon to be highly capable at cyber defense, to the point where it can autonomously find, validate and patch critical software vulnerabilities. A model that good at finding security holes could, in the wrong hands, help attackers too.
So Google is starting with defenders. It describes the rollout as phased, says it’s actively engaged in the U.S. government’s voluntary process for pre-release model access, and says it will gather feedback from early testers while refining guardrails before releasing Argon to developers, enterprises and consumers.
Gemini 4 Features and Capabilities
Everything below comes from Google’s own announcement unless noted otherwise. Independent testing is still limited because so few people have access.
A much larger output limit. Google raised the model’s output limit to 1 million tokens, up from the previous 64,000. Tokens are small chunks of text, so this means Argon can produce and reason through far longer pieces of work in one go. Google says this extra headroom lets the model think deeply across hundreds of thousands of tokens to solve hard problems in a single run.
Note that this is the output limit. Google hasn’t published Argon’s input context window (how much it can read at once) in the announcement.
Software engineering. Google says engineers use Argon for everything from routine debugging to large codebase migrations and algorithm design. Internal examples include Argon agents migrating C and C++ code to Rust across Google, scaling up to more than 800,000 lines for the Fuchsia Zircon kernel. Google notes those rewrites go through automated and manual audits before reaching production.
Enterprise knowledge work. Google says Argon leads on benchmarks covering finance, legal and tax work, and it highlights results on Harvey’s Legal Agent Benchmark (legal research and drafting) and Vals Finance Agent v2 (multi-step financial research).
Visual understanding. Google describes Argon as especially strong when work involves visuals: analyzing charts, picking out details in long videos, and acting on a series of documents.
Agents and tool use. Several of Google’s examples involve “Argon agents” carrying out tasks on their own. For instance, Google says a team of Argon agents analyzed data center telemetry and applied memory optimizations, freeing more than 300 TiB of memory.
Cybersecurity defense. Security firm Wiz is using Argon through its Scan for Good initiative. Google says the model found a critical vulnerability exposing sensitive personal information in hospital software used worldwide, one that earlier frontier models had missed.
Gemini 4 Release Date
Here’s where things stand:
- Announced: September 30, 2026
- Available now to: a set of trusted cyber defenders in the Fairwind Program, plus Google’s internal teams
- Next in line: paid API customers and Google AI Ultra subscribers
- General public release: no date announced
Google’s only timing language is “as soon as possible.” Android Authority noted that Google hasn’t given an ETA beyond “soon.” Any specific public release date you see online is not from Google.
Who Can Use Gemini 4 Right Now?
Only participants in Google’s Fairwind Program and Google’s own teams.
Fairwind is Google’s limited-access cyber defense program. Reporting describes it as a global program aimed at governments, critical infrastructure and core technology platforms, with more than 650 partners. Google hasn’t named which partners are testing Argon or explained how they were selected.
One unusual detail: Fairwind participants and Google’s internal teams get a version of Argon without cyber guardrails, so they can use its full defensive capabilities. The broader release will include safeguards.
If you use the Gemini app or AI Studio today, you won’t find Argon there yet.
Gemini 4 Pricing
Google has published API pricing, though it applies once Argon is available to paying API customers:
| Pricing | Rate |
|---|---|
| Introductory input | $2 per 1M tokens |
| Introductory output | $10 per 1M tokens |
| Cached input | 95% off the input price |
| Standard input (after intro) | $4 per 1M tokens |
| Standard output (after intro) | $20 per 1M tokens |
Google states the introductory rates first and, in a footnote, the standard $4 and $20 rates that apply after the introductory period ends. It hasn’t said how long the introductory period lasts.
For consumers, Google says Argon will come first to Google AI Ultra subscribers. It hasn’t announced whether Argon will reach free Gemini users or lower paid plans.
Gemini 4 vs Previous Gemini Models
Here’s how Argon fits into Google’s recent lineup:
Gemini 3
- Released: November 2025
- Role: the previous flagship generation
Gemini 3.1 Pro
- Released: February 2026
- Role: Google’s most recent “Pro” flagship before Argon
Gemini 3.5 Pro
- Status: reportedly cancelled, according to 9to5Google, with Google focusing on 3.8 Flash instead
Gemini 3.8 Flash and 3.8 Flash Cyber
- Released: September 2026
- Role: faster models, with a security-focused variant. Google says Argon makes large leaps in vulnerability discovery over 3.8 Flash Cyber.
Gemini 4 Argon
- Announced: September 30, 2026
- Role: first Gemini 4 frontier model, for complex long-running work
- Access: limited to Fairwind and Google’s internal teams for now
Gemini 4 vs ChatGPT and Claude
It’s too early for a fair head-to-head, because Argon isn’t broadly available and most comparison data so far comes from Google. Here’s what has been documented, by source.
What Google and reporting say about benchmarks
- Google says Argon is the leading model on the Vals Index, which weighs finance, coding, legal and tax work by each sector’s contribution to U.S. GDP.
- CNBC reported Argon ties OpenAI’s GPT-6 Astra and xAI’s Grok 4.7 on cybersecurity benchmarks, and ranks ahead of GPT-6 Astra and recent Anthropic models on the Vals Index.
- Axios reported Google says Argon outperforms GPT-6 Astra on several coding and knowledge-work benchmarks.
Where the picture is less clear
- Google’s claims don’t show Argon leading every test, and independent reviews are still limited.
- The Next Web, citing Bloomberg, reported some internal doubts at Google about the model’s performance; Google told Bloomberg it would be inaccurate to say Gemini 4 underperforms in areas such as coding.
Availability and deployment
- OpenAI’s and Anthropic’s current flagship models are available to their paying customers. Argon, for now, is limited to a vetted group.
- Google is making the cyber-unrestricted version available only to trusted defenders, while the general release will carry safeguards.
The practical takeaway: these are company-reported and early figures. Once Argon reaches API customers, independent testing will give a clearer picture.
Gemini 4 Benchmarks
All results below come from Google’s announcement.
| Benchmark (what it measures) | Argon result |
|---|---|
| DeepSWE v1.1 (long software engineering tasks) | 77.9%, a new high per Google |
| CWE-bench v1 (fixing security vulnerabilities) | 68%, tied for first |
| AutomationBench by Zapier (end-to-end business tasks) | 51.3%, ranked #1 |
| LVBench (long video understanding) | 91.7%, state of the art per Google |
Google also says Argon leads on the Vals Index, Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark, and on Gray Swan’s indirect prompt injection benchmark (a measure of resistance to hijacking attempts), without giving a single headline number for each in the text.
Two cautions: these are Google-reported results, and “tied for first” on CWE-bench means other models matched it.
Gemini 4 Safety: What Google Is Doing First
Google says it’s strengthening safeguards in four areas before a broad launch:
- Misuse: designing Argon to refuse requests that could help cyber, chemical, biological, radiological or nuclear attacks, while still allowing legitimate research.
- Prompt injection: hardening it against hidden malicious instructions in content the model reads. Google calls Argon its most resilient model yet on this front.
- Misalignment: monitoring the model’s reasoning and actions, and stopping it if it tries to go beyond what the user intended.
- Hardened systems: isolating and sealing test environments before high-risk training or evaluation.
The launch also came a day after Google and other AI companies joined a voluntary U.S. safety agreement covering pre-release evaluations, according to Tech Wire Asia.
Why Gemini 4 Matters
- For cybersecurity teams: Fairwind participants get early access to a model Google built specifically for finding and fixing vulnerabilities.
- For developers: once API access opens, the 1 million token output limit and published pricing give a clear sense of what large coding and agent projects might cost.
- For businesses: Google is pitching Argon directly at legal, finance and enterprise automation work, not just chat.
- For everyday Gemini users: nothing changes yet. AI Ultra subscribers are first in line when consumer access arrives.
What Happens Next?
Google has laid out the order, but not the timing: Fairwind cyber defenders first, then paid API customers and Google AI Ultra subscribers, then broader availability for developers, enterprises and consumers.
What’s still unconfirmed: the exact release dates, Argon’s input context window, how long introductory pricing lasts, whether free Gemini users will get access, and whether other Gemini 4 models will follow.
For now, Gemini 4 exists as Argon, it’s real, and it’s in very few hands. Expect wider access to arrive in stages once Google is satisfied with its safeguards.
Gemini 4 FAQ
What is Gemini 4?
Google’s next generation of Gemini AI models, developed by Google DeepMind. Gemini 4 Argon is the first model in it.
What is Gemini 4 Argon?
A frontier reasoning model built for long, complex work in software engineering, legal and financial tasks, and cybersecurity defense.
When was Gemini 4 announced?
September 30, 2026.
Is Gemini 4 available to the public?
No. It’s available only to trusted cyber defenders in Google’s Fairwind Program and Google’s internal teams.
When will Gemini 4 be available?
Google hasn’t given a date. Paid API customers and Google AI Ultra subscribers come next, followed by broader availability.
Can I use Gemini 4 for free?
Not currently. Google hasn’t announced whether free Gemini users will get access.
How much does Gemini 4 cost?
API pricing starts at an introductory $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. Consumer access will start with Google AI Ultra.
Is Gemini 4 available through the API?
Not yet. Paid API customers are next in line after the Fairwind rollout.
Is Gemini 4 better than ChatGPT or Claude?
There’s no settled answer. Google reports Argon leading several benchmarks, and CNBC reported it tying or beating OpenAI’s GPT-6 Astra on some. Most data so far comes from Google, and independent testing will matter once the model is widely available.
Sources
- Google Blog: Gemini 4 Argon announcement: model description, rollout plan, pricing, output limit, benchmarks, internal uses, safeguards.
- CNBC: competitor benchmark comparisons and comments from Tulsee Doshi.
- Axios: context on Google’s model lineup and benchmark claims versus GPT-6 Astra.
- 9to5Google: Gemini 3.5 Pro cancellation and naming change.
- Android Authority: no ETA beyond “soon.”
- The Next Web: Bloomberg-reported internal doubts and Google’s response.
- The Hack Academy: Fairwind program scope and undisclosed tester details.
- Tech Wire Asia: voluntary U.S. safety agreement timing.
- Wikipedia: Google Gemini: previous model release dates.
