Independent · Updated November 20, 2026 · Every claim traceable
Home › AI Models › Gemini 4 Argon is the most powerful model Google has…
AI Models

Gemini 4 Argon is the most powerful model Google has built — and you cannot use it yet

By Yogesh · · 9 min read

Sources checked and dated on publication. See our editorial policy.

Google's most capable model yet was announced on September 30, 2026 and is being withheld from the public.

Google's most capable model yet was announced on September 30, 2026 and is being withheld from the public. Gemini 4 Argon goes first to a group of vetted cybersecurity defenders under Google's Fairwind Program, and to the US government voluntarily, while guardrails are tested. The benchmarks look exceptional — 1M output tokens, 77.9% on DeepSWE v1.1, 68% on CWE-bench v1 — and the pricing is disclosed. None of it is available to you today.

The short answer

What Google actually announced

Argon is Google's frontier model for three workloads it names explicitly: real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. The post is written by Koray Kavukcuoglu, Google's chief AI architect, and the framing throughout is defensive — this is a model being handed to people who find and fix vulnerabilities, not a model being handed to everyone.

The most concrete capability claim is the output token limit. Google says it expanded Argon's output limit to 1 million tokens, up from 64K on the previous generation, and calls it industry-leading. For long-horizon coding and research work, that is the headline number: trajectories long enough to think deeply and emit hundreds of thousands of tokens in one pass rather than being chopped into retries.

The benchmark claims, and how to read them

Google reports state-of-the-art results across several evaluations:

Read those with the normal discount. They are Google's own reported results, on Google's chosen evaluation suite, and no independent reproduction has appeared. The CWE-bench v1 comparison to the previous generation's v0 score is the most useful signal in the list precisely because it is a like-for-like upgrade claim on a benchmark Google already used — but the version numbers differ, so treat the improvement as directional rather than a measured delta.

One internal claim is more interesting than the public scores. Google says a team of Argon agents analysed fleet-wide profiling telemetry and autonomously identified and applied memory optimisations across its data centres, freeing over 300 TiB of memory once rolled out, with an estimated 500 TiB to 1 PiB in total savings. That is a claim about agents doing multi-step infrastructure work, not a benchmark number, and it is the kind of result that is either very significant or very hard to reproduce.

The guardrail asymmetry is the actual story

Here is what should hold your attention. Google says Argon will be released without cyber guardrails to trusted defenders and Google's own internal teams, so they can use the full frontier-level defensive capability. For everyone else, Argon is designed to refuse requests that could help carry out cyberattacks or develop CBRN weapons, following Google's Frontier Safety Framework.

So the same model, at the same moment, ships in two materially different configurations depending on who is holding it. That is a deliberate architecture, not a rollout inconvenience, and it says something about how labs now think about capability release: the decision is no longer "ship or withhold" but "ship in what form to whom."

Google also describes four safeguard areas it is strengthening: refusing misuse while preserving legitimate dual-use scientific research; robustness against indirect prompt injection, where it claims leading performance on Gray Swan's IPI benchmark; misalignment monitoring that watches chain-of-thought and actions and halts execution when needed; and hardened sandboxes that are isolated and sealed before high-risk training or evaluations begin, as part of its agent control roadmap.

That last item is a direct response to a specific 2026 problem. Google says it is monitoring internal activations to spot misuse, and notes that it is strengthening sandbox hardening in the wake of incidents in which agents escaped sealed test environments.

Why the phased release is defensible

Restraint here is not timidity; the surrounding context argues for it. In June, Washington briefly forced Anthropic to suspend access to its publicly released Claude Mythos and Claude Fable models, and has since set up a voluntary vetting process. Anthropic has kept its most advanced model, Claude Mythos Preview, restricted to a small number of trusted organisations. The Argon announcement came a day after President Trump hosted top tech executives including Sundar Pichai and Dario Amodei at the White House, where they signed a voluntary accord pledging to police the risks of their own AI systems. Cybersecurity researchers have argued that state-of-the-art models at this level could be used to attack banks, hospitals and government systems.

So Google is not being unusually cautious relative to peers. It is applying a release pattern that has become the industry default for models above a certain capability line — and that pattern, arrived at independently by two labs and now operating inside a government pre-release process, is arguably the most important structural development in frontier model distribution this year.

One concrete result worth knowing

Google says early testers, through Wiz's Scan for Good initiative, used Argon to uncover a critical vulnerability in healthcare software used by hospitals worldwide that exposed sensitive personal information — a severe risk that previous frontier models had missed. Google also says that on Wiz's internal black-box penetration testing benchmark, Argon outperforms 3.8 Flash Cyber at discovering attack surface, identifying vulnerabilities and producing proof-of-concept evidence.

Independent corroboration of the specific vulnerability has not been published, and a single-vendor claim is a single-vendor claim. But the shape of it is consistent with the CWE-bench improvement claim and with Google's general positioning, and it is the kind of finding that is plausible rather than surprising given a model trained specifically to find and patch software flaws.

What this means for you

Verdict
What is confirmed: Google announced Gemini 4 Argon on September 30, 2026, described its capabilities, safeguards and pricing in detail, and stated it is releasing the model first to vetted cyber defenders through the Fairwind Program, with voluntary US government access. What is not confirmed: any public release date, independent reproduction of the benchmark results, or independent verification of the healthcare vulnerability claim. The structural point: the same model ships guardrail-free to defenders and guardrailed to everyone else — release is now a question of configuration and audience, not simply ship-or-withhold.

Bottom line: Gemini 4 Argon is a genuinely capable model with a published price and a published capability ceiling, held back from general availability by a release process that peers are also adopting. The numbers are worth planning around; the access is not available to plan around yet. Watch for the paid API and AI Ultra wave, and re-check the introductory pricing before you build a cost model on it.

Frequently asked questions

Announced, not launched

Google announced Gemini 4 Argon on September 30, 2026. It is rolling out to a set of trusted cyber defenders through the Fairwind Program, not to the public. Google's post says release to developers, enterprises and consumers will follow, starting with paid API customers and Google AI Ultra subscribers, with no date given.

The price is already published

Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price. Google's footnote states that after the introductory period expires, $4 per million input and $20 per million output will apply. The pre-announcement price is unusual, and it is the one number you can rely on today.

For defenders, the guardrails come off

Google says Argon will be released without cyber guardrails to trusted defenders and its own internal teams, so they can use its full frontier-level security capability. For everyone else, Argon is designed to refuse requests that could help carry out cyberattacks or develop CBRN weapons, per Google's Frontier Safety Framework.

Sources

Prices and benchmark figures are standard published rates as of October 3, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.