Gemini 4 Argon: Google's new frontier model for long coding, knowledge and cyber defense work

1 week ago

Gemini 4 Argon is Google's new frontier model for long coding, knowledge work and cyber defense tasks. It is rolling out first to trusted cyber defenders through the Fairwind Program, with paid API customers and Google AI Ultra subscribers next, at an introductory $2 in and $10 out per million tokens. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ https://deepmind.google/models/gemini/ https://deepmind.google/fairwind-program/ https://ai.google.dev/gemini-api/docs/models

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Gemini 4 Argon, Google's new frontier model
  2. 0:10Built for long, multi-step work
  3. 0:21Output limit rises to 1M tokens
  4. 0:33What Google reports
  5. 0:35Top score on long software tasks
  6. 0:49A video decoder rewritten in safe Rust
  7. 1:04Leading on finance, legal and tax work
  8. 1:20Charts, documents and long videos
  9. 1:33Finds, validates and patches vulnerabilities
  10. 1:45Wiz found a hospital software flaw with it
  11. 1:59Who gets it
  12. 2:01Fairwind partners only, for now
  13. 2:15Paid API customers and Ultra subscribers next
  14. 2:27$2 in, $10 out, at an introductory price
  15. 2:40Not in the Gemini API yet
Show transcript

Gemini 4 Argon, Google's new frontier model

010203What it is
blog.google · Gemini 4 Argon · Sep 30, 2026
Gemini

Gemini 4 Argon is Google's new frontier model

Rolling out to trusted cyber defenders first

Google has announced Gemini 4 Argon, its new frontier model. It is rolling out first to trusted cyber defenders, through Google's Fairwind Program.

Built for long, multi-step work

Gemini 4 Argon
010203What it is
blog.google · Gemini 4 Argon · Sep 30, 2026

Argon is built for long, multi-step work

Software engineeringdebugging to codebase migrations
Knowledge worklegal and finance
Cyber defensefinding and fixing vulnerabilities

Google built it to keep reasoning through long, complex work. That means real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense.

Output limit rises to 1M tokens

Gemini 4 Argon
010203What it is
blog.google · Gemini 4 Argon · Sep 30, 2026

One answer can now run to 1M output tokens

Up from 64K tokens before.

For builders, one response can now run to a million output tokens. The previous limit was 64,000. Google says the headroom lets it solve a hard problem in one pass.

What Google reports

Gemini 4 Argon
010203What Google reports
02
What Google reports

What Google reports.

Top score on long software tasks

Gemini 4 Argon
010203What Google reports
blog.google · Gemini 4 Argon · Sep 30, 2026 · deepmind.google · Gemini 4 Argon

Google reports a new best score on long, real-world coding tasks

Google's own table puts it behind rivals on Terminal-bench and FrontierSWE.

Google reports Argon sets a new state of the art on DeepSWE, a test of long, real-world software tasks. Google's own results table also shows rivals ahead on Terminal-bench and FrontierSWE.

A video decoder rewritten in safe Rust

Gemini 4 Argon
010203What Google reports
blog.google · Gemini 4 Argon · Sep 30, 2026

Argon agents made Google's Rust video decoder 2.7× faster

libgav1: 32K lines of SIMD code replaced with safe Rust, identical video output.

Inside Google, Argon agents rewrote the libgav1 video decoder's Rust port. They replaced about 32,000 lines of SIMD code with safe Rust. Google says it now runs 2.7 times faster, with identical output.

Leading on finance, legal and tax work

Gemini 4 Argon
010203What Google reports
blog.google · Gemini 4 Argon · Sep 30, 2026

Google reports Argon leads on finance, legal and tax work

Vals Index: tasks weighted by each sector's share of U.S. GDP.

Beyond code, Google reports Argon leads the Vals Index. It mixes finance, coding, legal and tax work, each weighted by its share of the U.S. economy. Google also reports first place on Zapier's business automation test.

Charts, documents and long videos

Gemini 4 Argon
010203What Google reports
blog.google · Gemini 4 Argon · Sep 30, 2026

It reads charts, document sets and long videos

Google reports a new best score for long-video understanding.

Argon also works with images and video. Google says it can analyze charts, pick details out of long videos, and act on a set of documents. Google reports a new best score on long-video understanding.

Finds, validates and patches vulnerabilities

Gemini 4 Argon
010203What Google reports
blog.google · Gemini 4 Argon · Sep 30, 2026

Argon can find, validate and patch software vulnerabilities

Trusted defenders get it without cyber guardrails.

Google trained Argon for cyber defense. It can find, validate and patch critical vulnerabilities on its own. Trusted defenders and Google's own teams get it without cyber guardrails.

Wiz found a hospital software flaw with it

Gemini 4 Argon
010203What Google reports
blog.google · Gemini 4 Argon · Sep 30, 2026

Wiz used Argon to find a critical flaw in hospital software

Earlier frontier models had missed it.

The security company Wiz uses it in Scan for Good, its free program protecting public infrastructure. Google says Argon found a critical flaw in hospital software worldwide that earlier frontier models had missed.

Who gets it

Gemini 4 Argon
010203Who gets it
03
Who gets it

Who gets it, and what it costs.

Fairwind partners only, for now

Gemini 4 Argon
010203Who gets it
deepmind.google · Fairwind Program

Today, only approved Fairwind partners can use Argon

Security teams onlycyber, incident response, pen testing
Zero data retentionon Gemini Enterprise

Right now, only approved Fairwind partners get Argon. They may give it only to internal security, incident response and penetration testing teams. On Gemini Enterprise, it supports zero data retention.

Paid API customers and Ultra subscribers next

Gemini 4 Argon
010203Who gets it
blog.google · Gemini 4 Argon · Sep 30, 2026

Paid API customers and Google AI Ultra subscribers come next

Then developers, enterprises and consumers, "as soon as possible".

Google is widening access gradually while it tests guardrails. Paid Gemini API customers and Google AI Ultra subscribers come next. Google gives no date, only "as soon as possible".

$2 in, $10 out, at an introductory price

Gemini 4 Argon
010203Who gets it
blog.google · Gemini 4 Argon · Sep 30, 2026

Introductory price: $2 in, $10 out per million tokens

Cached input95% off the input price
After the intro period$4 in, $20 out

The introductory price is 2 dollars per million input tokens and 10 dollars per million output tokens. Cached input is 95 percent off. Afterwards, it doubles to 4 and 20 dollars.

Not in the Gemini API yet

Gemini 4 Argon
010203Who gets it
ai.google.dev · Gemini API models · blog.google · Gemini 4 Argon · Sep 30, 2026

Argon has no Gemini API model ID yet

Announcement: blog.google › Gemini 4 Argon · Watch: ai.google.dev › Gemini models

For developers today, Argon is missing from the Gemini API models page, with no model ID to call. Watch that page for it. That's Gemini 4 Argon, Google's new frontier model.