Gemini 4 Argon: the first frontier model since the accord goes to cyber defenders first, with the guardrails off

On September 29, 2026 Sundar Pichai sat in the White House and signed a one-page pledge, alongside the chief executives of Anthropic, Meta, OpenAI, xAI and Nvidia, promising that the companies racing toward superintelligence would monitor their own models and open them to outside auditors. On September 30 Google announced Gemini 4 Argon, the first frontier model released by any signatory since the ink dried. The release is the earliest evidence of what the accord means in practice, and the answer so far is: not much changes, except who gets the model first.

Argon is not a public product yet. It is available to members of Google's Fairwind Program, a group of cyber defenders that Google vets, and to Google's own security teams. Paid API customers and Google AI Ultra subscribers come next, with no date. Everyone else waits.

What Google says the model does

Koray Kavukcuoglu, who has run Google DeepMind since Demis Hassabis moved to Alphabet's chairmanship in August, wrote that Argon "delivers frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense." The headline change is the output limit: one million tokens, up from 64,000 on the previous generation, which Google says gives the model room to reason through a long task in a single pass rather than in stitched-together steps.

The numbers Google published are specific. On DeepSWE v1.1, a long-horizon software engineering benchmark, Argon scores 77.9%, against 74.2% for Anthropic's Claude Opus 5.5 and 74.1% for OpenAI's GPT-6 Astra in Google's table. On CWE-bench v1, which measures fixing security vulnerabilities, Argon ties Astra at 68%, with Opus one point behind. On AutomationBench Google claims first place at 51.3%, and on LVBench, a long video test, 91.7%. On real-world vulnerability discovery, the capability that makes the cyber-defender release coherent, Google reports 85.8%, up from 71% for its Gemini 3.8 Flash Cyber model.

The table is not uniform. On Terminal-Bench 4.0, a coding benchmark that rewards sustained agentic work, Opus 5.5 leads at 66.4% to Argon's 57.4%. Reuters, as carried by Tech Wire Asia, summarized Google's own evaluations as putting Argon "ahead of OpenAI's Astra and Anthropic's Opus on some cybersecurity and other benchmarks, while it trails competing models on some coding tests." An independent test by Artificial Analysis, reported by Trending Topics, puts Argon level with Astra and about five points behind Opus 5.5 at the top setting. Google's claim to lead the frontier is therefore true on its own table, contested on independent ones, and untestable by anyone outside Fairwind for now. Our sister title Super Intelligence News UK has the full benchmark table and the UK access picture.

Who gets it, and who gets it without guardrails

Fairwind is the part of this release that matters for the race. Google launched the program on September 2, 2026 to give what it calls "high-priority defenders (like governments, healthcare providers, and telecommunications services) early access to advanced models that help them build better defenses, before new threats arrive." More than 650 partners are in it. The access rules for Argon are tighter than for a normal API key: only internal cybersecurity, incident response or penetration testing teams may use it, organizations must run phishing-resistant multi-factor authentication, they must track which employees use the model, and access may not be shared, resold or redistributed.

Inside that circle, Google is going further. The Hacker News reported, and Tech Wire Asia confirmed from Google's materials, that trusted Fairwind users and Google's internal security teams receive a version of Argon without its cyber guardrails, so that defenders can use the model's full offensive capability to find holes before attackers do. Google says Argon already found a previously unknown critical vulnerability in a piece of healthcare software that exposed personal information; it has not named the software.

For everyone else the guardrails are the product. Google says the public model refuses help with cyber attacks and with chemical, biological, radiological and nuclear weapons, is its most resistant model yet to indirect prompt injection, and tops Gray Swan's injection benchmark. It is also "deploying misalignment mitigations that monitor Argon's chain-of-thought and actions and stop execution when necessary," and it seals the sandboxes used for high-risk evaluations before testing begins. That last line reads differently after OpenAI's two sandbox escapes this summer, the second of which forced a training pause in late September; we keep the record on our rogue AI incidents log.

What the accord did and did not change

The accord Pichai signed commits the six companies to internal monitoring of capability and alignment, a dedicated internal team, independent external auditors, an independent board committee and regular meetings between the firms, according to Nextgov's report on the signing. It carries no enforcement. Vice President JD Vance put the administration's view plainly at the ceremony: "The solution to some of the AI risks is for you guys to take the risk seriously, not to come to the government for a regulatory regime." We explain what the accord commits the companies to and what the renaming order does and does not do elsewhere on this site.

Measured against that pledge, the Argon release looks like business as usual with better paperwork. Google says it is "actively engaged in the U.S. government's voluntary process for pre-release model access," which is the Center for AI Standards and Innovation testing that five labs had already agreed to in May. No external auditor is named. No board committee is mentioned. The staged release to defenders predates the accord: Fairwind opened four weeks before Pichai signed.

What is new is the price signal. Google's introductory rate is $2 per million input tokens and $10 per million output tokens, with cached input 95% off, rising to $4 and $20 after the introductory period. That is a frontier model priced to be used at scale by enterprises, not a research preview, and it tells you how Google expects to win: not by being first to superintelligence, but by being the cheapest serious model on the desk of every bank, law firm and security operations center.

What to watch

Three things decide whether this release moves the race or just the leaderboard. First, when paid API access opens, independent evaluators will be able to run the model, and the gap between Google's table and Artificial Analysis's numbers will either close or widen. Second, whether the guardrail-free build stays inside Fairwind; a model built to find vulnerabilities is also a model built to exploit them, and 650 partners is a lot of doors. Third, whether any of the accord's promised auditors appear with a name attached. Our position: Argon is a strong model and a strong enterprise price, and the defender-first rollout is a reasonable way to ship a capable cyber tool. But the release is the first test of the accord, and on the evidence published so far it has been met with the same voluntary measures the labs were already taking. We will update this piece when the API opens and when any independent evaluation of Argon is published.