Google Shipped Gemini 4 Argon With a Build That Has No Cyber Guardrails. The Price Expires in Seven Weeks

Google released Gemini 4 Argon at 16:00 UTC on 13 October, and the launch page leads with capability: a frontier model for "deep work" in coding, finance, legal, cybersecurity and science. Two details further down the page matter more than any of the scores. The price on the pricing line expires in seven weeks. And one version of the model ships to vetted outside parties with its cybersecurity guardrails removed.

The price is an introductory price

Google lists Argon on the Gemini API at $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95%. The footnote says those rates apply until 30 November 2026, after which the public-access model costs $4 in and $20 out.

That second row is the real price, and it is the same $4 and $20 Meta set for Muse Wonder on 12 October. So Argon is not priced below the frontier; it is priced at the frontier with a seven-week discount attached. The discount is well targeted — it covers exactly the window in which developers port agent harnesses from one model to another, and switching costs are what keep them there afterwards. Anyone building a cost model on the launch-day number is building it on a number that doubles before the quarter ends.

A version with the cyber guardrails taken off

The more consequential decision is the Fairwind Program. Google gives vetted cybersecurity defenders a build of Argon without its cyber guardrails, and Reuters reports the company is now expanding that from the select partners it started with in September to hundreds more.

The reasoning is sound and the labs have been converging on it. A model that refuses to discuss exploitation is useless to the people doing vulnerability research, incident response and red-teaming, and refusal does not stop a determined attacker who can fine-tune an open-weights model instead. Defenders face a volume problem that general-purpose models could genuinely help with.

The difficulty is that "hundreds of partners" is a different security posture from "select partners". Every added organisation is another place the vetting has to hold, another set of credentials worth stealing, and another insider who could misuse the access. Google has not published the vetting criteria, the review cadence, or what happens when a partner is compromised. For a capability tier this sensitive, those documents are the ones that would let anyone outside Google judge the programme, and they do not exist publicly.

Every capability number is Google's or a private eval

Argon is reported at #1 on the Vals Index across finance, coding, legal and tax work; 87% on the ERQA embodied-reasoning benchmark; and 57% on RareBench and 52.2% on LabBench, the latter two from Gamow Labs' private evaluations. Google also says Argon leads the Gray Swan indirect prompt-injection benchmark with attack success cut to 0.7% across 15 adaptive attempts, and that shortcut attempts on coding benchmarks fell to 2%.

The productivity figures are internal: teams using Argon in Antigravity raised agentic code submissions by more than 35% and cut code rollbacks by nearly a third. Those are Google's measurements of Google's customers, reported by Google.

None of this is unusual, and none of it is independently reproduced yet. The prompt-injection result is the one worth watching, because it is the claim with the clearest defensive value and the one a third party can most readily test. Until someone does, the verifiable facts about this launch are the price schedule and the distribution list.

Why the launch carried this much weight

Argon arrives after a bad year for Google's model cadence. In September the company said it would not ship Gemini 3.5 Pro at all, a model Sundar Pichai had earlier said would land in June. DeepMind was reorganised after co-founder Demis Hassabis stepped aside, and several leaders of the Gemini effort left. Reuters frames the release as Google trying to recover ground against OpenAI and Anthropic after months of delay.

The distribution answers that framing. Argon is in the Gemini API, AI Studio, Android Studio, Gemini for business and Google Antigravity; AI Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search. Google named Deloitte, LexisNexis, Ambience Healthcare and Isomorphic Labs as early customers, and is offering 10,000 scientists six months on Antigravity's top usage tier. Koray Kavukcuoglu's post says the rest of the Gemini 4 family follows.

That is a company shipping breadth rather than a single flagship number, which is the right move when the previous flagship was cancelled. Whether the model justifies the parity price is a question its own benchmarks cannot settle, and the clock on the discount means buyers have until 30 November to find out.

Sources: Google: Gemini 4 Argon: Built to solve your hardest problems · Reuters: Google launches flagship AI model Gemini 4 Argon for developers, businesses · Meta developer docs: Muse Wonder pricing