Google's Gemini 4 Argon Rolls Out to Cyber Defenders First
TL;DR
- No independent lab has reproduced a single Argon benchmark score; every published result originates from within the Fairwind pre-release cohort.
- Argon trails GPT-6 Astra by roughly 10 points on FrontierSWE terminal-driven tasks, directly undercutting the broad coding-lead framing.
- At $2/$10 per million input/output tokens, Argon is priced roughly 5x below GPT-6 Astra, making cost efficiency the more defensible competitive argument.
Google released Gemini 4 Argon on September 30, 2026, and the first users are not paying customers but trusted cyber defenders inside its Fairwind Program. Paid API access and Google AI Ultra subscribers come after, and the company says it is engaged in the U.S. government's voluntary pre-release model access process.
The announcement, written by SVP of Google DeepMind and Chief AI Architect Koray Kavukcuoglu, leans on code and security numbers. Argon posts 77.9% on DeepSWE v1.1, ties for first at 68% on CWE-bench v1 for vulnerability remediation, and lands at 51.3% on AutomationBench for end-to-end business tasks. Output length now stretches to one million tokens, up from 64K on the prior generation. Introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 after the intro window. Cybersecurity firm Wiz is named as an early user through its Scan for Good initiative, one more entry in what has already been a heavy quarter of AI cybersecurity coverage on our tracker.
Bloomberg reported the same day that some Google employees with direct access to the model say Argon performs less well on real coding tasks than the benchmarks suggest. Google told Bloomberg the characterization is inaccurate. One employee said there is "large consensus" inside the company that the model is at the frontier.
Kavukcuoglu's own line in the post: "Argon is fundamentally changing the way we work and build at Google." That is a claim about internal use, published the same week Bloomberg's sources describe the internal picture as less uniform. A broader release date for the paid tiers was not announced.
What others are reporting
-
Cyber Kendra Read →
Cybersecurity-specialist read: confirms Fairwind launched without standard safety guardrails and surfaces a 15% hallucination rate from independent testing absent from Google disclosures.
Argon is a well-rounded model that has frontier capabilities across several domains — Tulsee Doshi, head of Gemini products at Google DeepMind
-
DataCamp Read →
Critical perspective: notes no outside party has reproduced any score, details terminal-task underperformance versus Astra, and flags no published API model ID at launch.
Argon can produce up to 1M output tokens in one generation, against 64K on the previous generation.
-
OfficeChai Read →
Grounds the launch in pricing ($2/$10 per M tokens) and internal production results including a 40% quantum algorithm gain and an 800K-line Rust kernel migration.
Argon is a model built to sustain deep reasoning across long-horizon workflows, with strengths in real-world software engineering and cybersecurity defense.
-
Basic Tutorials Read →
Confirms the 13-of-18 benchmark lead but notes Astra still leads on specific coding and science categories; adds the 5x pricing gap as the sharper competitive differentiator.
Argon outperforms OpenAI's GPT-6 Astra in 13 out of 18 published benchmarks, including programming, automation tasks, and understanding long videos.
Shared on Bluesky by 6 AI experts (top 5 by trust)
Originally reported by blog.google
Read the original article →Original headline: Google Launches Gemini 4 Argon to Cybersecurity Partners First, Claims Coding Lead Over GPT-6 Astra