Gemini 4 Argon

Learning Objectives

• Understand what Gemini 4 Argon is and why Google is releasing it to cyber defenders first
• Read Google's benchmark claims for what they are: vendor-reported and not yet independently reproduced
• Know who can use Argon today, what it will cost, and what to use in the meantime

What Is Gemini 4 Argon?

Gemini 4 Argon is Google DeepMind's next frontier model, announced on September 30, 2026 by Koray Kavukcuoglu, Google's chief AI architect. It is the first new flagship since Gemini 3.1 Pro in February. Google had promised a Gemini 3.5 Pro for the summer and shipped a run of Flash models instead; Argon jumps the version number to 4 rather than closing that gap.

Google positions Argon for long, multi-step professional work: real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. The most concrete change is length. Argon's output limit is 1 million tokens, up from 64,000 in earlier Gemini models, so a single run can produce hundreds of thousands of tokens of reasoning and output rather than stopping and being restarted.

Why Cyber Defenders Go First

Google says releasing capabilities at this level safely requires a phased approach, and that it is taking part in the US government's voluntary process for testing models before release. Argon was trained to find, validate and patch critical software vulnerabilities on its own, and Google says it is giving the model to trusted defenders and its own teams without its usual cyber guardrails, so they get the full capability. That is the reason for the gate: the same skill that patches a flaw can find one to exploit.

The early example Google cites comes from Wiz, which is using Argon in its free Scan for Good program for critical public infrastructure. Google says the model found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, one that earlier frontier models had missed. Google did not name the software or publish details, so treat it as a claim rather than a demonstration.

Before a wider release, Google says it is strengthening four kinds of safeguard. It is improving refusals for cyber and chemical, biological, radiological and nuclear misuse while trying to keep legitimate research working. It is hardening the model against prompt injection, where it says Argon leads Gray Swan's indirect prompt injection benchmark. It runs a monitor that reads the model's chain of thought and actions and stops a run that goes beyond what the user intended. And it is sealing its testing sandboxes before high-risk training and evaluation begin. Google also urges other labs to keep model reasoning readable, so that monitors like this one keep working.

Benchmarks

Every figure below is reported by Google. None has been independently reproduced, and Google published them without the full evaluation settings.

Ars Technica reports that the DeepSWE score is ahead of GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Google also names leading results on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark without giving the numbers.

Google's examples of internal use are more specific than the benchmarks. It says a team of Argon agents analyzed memory profiling across Google's data centers and freed more than 300 tebibytes of memory, that Argon agents are migrating C and C++ code to Rust at scales up to the 800,000-line Fuchsia Zircon kernel, and that on Google's open-source video decoder the result runs 2.7 times faster than the earlier Rust port with identical output. Google adds that these rewrites still go through automated and manual auditing before reaching production.

Pricing & Access

The introductory API price is low for a frontier model. GPT-6 Astra lists at 10 dollars per million input tokens and 50 dollars per million output. But a price for a model you cannot call is a statement of intent, and Google has not said how long the introductory rate will last.

Strengths

• Output length: a 1 million token output limit lets one run carry a long migration, report or analysis without being broken into pieces
• Long-horizon coding and professional work: Google reports leading scores on DeepSWE, the Vals Index and AutomationBench
• Cyber defense by design: trained to find, validate and patch vulnerabilities, with a gated release that puts that skill with defenders first
• Readable reasoning as a safety tool: Google monitors Argon's chain of thought and argues publicly that other labs should keep theirs readable too
• Low announced price: the introductory API rate sits well below the closed frontier models it is compared with

Limitations & Considerations

• Not generally available: only Fairwind members can use it, and there is no date for anyone else
• Every benchmark is vendor-reported: no independent lab has reproduced Google's numbers, and the most striking claim, the Wiz hospital-software flaw, came with no details
• Released without cyber guardrails to defenders: that is a deliberate choice for vetted users, and it is also why access is restricted
• No context-window figure published: Google announced the output limit, not the input limit, so do not assume the two match
• Pricing will change: the introductory rate doubles to 4 dollars and 20 dollars per million tokens later, on a schedule Google has not set

Related Tools

• Gemini 3.1 Pro — the previous Pro-tier flagship and the one you can use today for deep reasoning
• Gemini Flash — Google's shipping workhorse line, including the Fairwind-gated 3.8 Flash Cyber
• Gemini — Google's consumer app, where Argon will reach Google AI Ultra subscribers first
• GPT-6 Astra — OpenAI's flagship, generally available since September 3, 2026, and the model Google benchmarks Argon against
• Wiz Cloud Security — the security company using Argon in its Scan for Good program

Lineage

• Gemini 4 Argon (announced September 30, 2026) — frontier model, Fairwind-only at launch; the page above
• Gemini 3.8 Flash (September 2, 2026) — the shipping workhorse, with a Fairwind-gated Cyber variant
• Gemini 3.5 Pro (announced May 2026) — promised for the summer and never released
• Gemini 3.1 Pro (February 2026) — the last Pro-tier model to ship

Key Takeaways

• Gemini 4 Argon is Google DeepMind's new frontier model and its first new flagship since Gemini 3.1 Pro in February 2026, jumping past a Gemini 3.5 Pro that never shipped
• Access at launch is limited to vetted security teams in the Fairwind Program; paid API customers and Google AI Ultra subscribers come next, with no date
• Google gives defenders the model without its usual cyber guardrails, which is why the release is gated rather than open
• Google reports 77.9 percent on DeepSWE and first place on the Vals Index and AutomationBench, but every number is vendor-reported and none has been reproduced
• The output limit rises to 1 million tokens from 64,000, the change most likely to matter for long migrations and reports
• The announced API price starts at 2 dollars per million input tokens and 10 dollars per million output, rising to 4 dollars and 20 dollars after an introductory period