Pull to refresh
Logo
Google launches Gemini 4 Argon, its most advanced AI model

Google launches Gemini 4 Argon, its most advanced AI model

New Capabilities

Flagship targets coding, cybersecurity, and enterprise work, but stays limited to vetted government and defense testers for now

Today: Argon restricted access begins

Overview

Updated 1 hour ago

Google spent 2026 shipping smaller models while OpenAI and Anthropic held the frontier. On September 30 it answered with Gemini 4 Argon, a flagship built for long, multi-step reasoning across coding, finance, legal work, and cybersecurity, restricted for now to vetted cyber defenders and U.S. government testers.

Benchmarking firm Artificial Analysis says Argon matches OpenAI's GPT-6 Astra on a composite intelligence index at 60% of the cost per task. The model also outputs up to 1 million tokens per response, sixteen times its predecessor's ceiling, so it can tackle long problems in a single pass.

Why it matters

Google can now match OpenAI and Anthropic at the frontier and undercut on price, giving builders a real third option.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

$2 / $10
Introductory API price per million tokens (input/output)
Cached input tokens cost 95% less, per Google's rollout pricing.
1M
Maximum output tokens per response
Up from the previous 64,000-token limit, an industry-leading ceiling.
77.9%
DeepSWE v1.1 software engineering benchmark
Measures real-world long-horizon coding tasks; Google calls it a new state of the art.
91.7%
LVBench long-video understanding score
Google says Argon leads the benchmark for understanding long videos.
68%
CWE-bench v1 cybersecurity remediation
Ties for first place on the security vulnerability-fixing benchmark.
60%
Cost per task vs. OpenAI GPT-6 Astra
Per Artificial Analysis, matches Astra's Intelligence Index at 60% of cost per task.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

Sign up to generate historical perspectives on this story.

People Involved

Organizations Involved

Timeline

June 2026 October 2026

4 events Latest: Today
Tap a bar to jump to that date
  1. Argon restricted access begins

    Today Product launch

    Initial cohort of cyber defenders and government testers gets access through the Fairwind Program.

  2. Google unveils Gemini 4 Argon

    Product launch

    Flagship announced and limited to vetted cyber defenders and U.S. government testers.

  3. Google rolls out smaller Flash models

    Release

    Lighter, lower-cost Flash models release through the summer while the company prepares its flagship.

  4. Google ships Gemini 3.5 Pro

    Release

    Mid-cycle upgrade after Gemini 3 put Google back at the frontier of the AI model race.

Scenarios

1

Google opens Argon to API customers and AI Ultra subscribers

Likely Resolves by Q2 2027

Discussed by: Google's own announcement; Ars Technica and InfoWorld on the restricted release

Google completes the U.S. government pre-release safety review and the Fairwind pilot, then flips Argon on for paid developers and Google AI Ultra subscribers. Pricing is already set at $2 per million input tokens and $10 per million output. The main question is when, not whether.

2

A rival model tops Argon's benchmark scores

Possible Resolves by Q2 2027

Discussed by: Artificial Analysis, the benchmarking firm that tracks model intelligence indices

OpenAI ships a successor to GPT-6 Astra or Anthropic releases the next Opus that posts higher scores on DeepSWE, Vals Index, or the Artificial Analysis Intelligence Index. With the model race moving this fast, Argon's lead may be brief.

3

Security incident or misuse tied to Argon's guardrail-free release

Unlikely Resolves by Q2 2027

Discussed by: SecurityWeek and Google's Frontier Safety Framework

Google released Argon to cyber defenders without cyber guardrails so it can defend at full capability. If those capabilities are misused, or a model-suggested patch contains a flaw, it could produce a notable security incident. Google says it monitors internal activations and chain-of-thought to catch misuse.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

February 2023

Google Bard stumble (February 2023)

Google rushed to answer OpenAI's ChatGPT with Bard, a chat AI. A demo showed Bard incorrectly claiming the James Webb Space Telescope took the first images of an exoplanet. Alphabet's market value dropped roughly $100 billion in the days after the error.

Then

Google pledged more careful testing and pulled back from aggressive AI launches.

Now

Google adopted staged, evaluation-heavy releases for its flagship models.

Why this matters now

Argon's phased, restricted release comes straight from that lesson: Google now withholds its best model until it is red-teamed and vetted.

March 2023

OpenAI's staged GPT-4 rollout (March 2023)

OpenAI released GPT-4 gradually, first to paid ChatGPT Plus subscribers and a handful of developers, after months of safety evaluation and red-teaming. Access was scarce for weeks.

Then

Scarce access built demand and let OpenAI monitor real-world use.

Now

Staged, safety-gated launches became the norm at frontier AI labs.

Why this matters now

Argon follows the same template: trusted testers first, broader access later, with a documented safety framework.

Sources

(11)