Gemini 4 Argon: Google leads 13 of 18 benchmarks, but only cyber defenders can use it

Google DeepMind launched Gemini 4 Argon on 30 September 2026, claims the lead on 13 of 18 benchmarks it disclosed, and then handed it to almost nobody. The model is rolling out first to trusted cyber defenders through Google's Fairwind Program, with paid API customers and Google AI Ultra users to follow. If you are asking what Gemini 4 Argon is and whether you can use it, the answer today is: read the benchmarks, but wait.

What Google announced

Koray Kavukcuoglu, Google DeepMind's senior vice president and chief AI architect, introduced Argon as the company's model for long-horizon work: real-world software engineering, legal and financial knowledge work, and cyber defence. The headline specification is output length. The single-answer limit rises from 64,000 tokens to 1 million, which lets the model plan and execute many steps inside one task.

Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at a 95% discount. Standard pricing of $4 and $20 follows after the introductory period, and the timing of that change has not been announced.

The benchmark claims

Google's own figures, as reported by VentureBeat and DataCamp, put Argon ahead on several tests that map to paid work. It scores 77.9% on DeepSWE v1.1, a real-world software engineering test, against 74.1% for OpenAI's GPT-6 Astra and 74.2% for Anthropic's Claude Opus 5.5. On AutomationBench it scores 51.3%. On Harvey's legal agent test it scores 19.6% against 5.4% for GPT-6 Astra. It ties for first on CWE-bench v1, a vulnerability remediation test, at 68%.

The picture is not one-sided. VentureBeat reports GPT-6 Astra still leads on FrontierSWE v2, 65.5% to 55.0%, and Claude Opus leads on Terminal-Bench 4.0, 66.4% to 57.4%. DataCamp adds that Argon trails Astra by more than 10 points on Terminal-Bench Science 0.1. In plain terms, Argon looks strongest on long documents and office-style agent work, and weaker on terminal-driven science tasks.

The caveat that matters most is DataCamp's: no third party has independently verified the published benchmarks. These are vendor numbers from a launch day. Treat them as a claim to be tested.

Why the access model is the story

A frontier lab withholding a flagship from the public for safety reasons is no longer unusual. What is new here is that Google says it will release Argon to trusted defenders "without cyber guardrails", in its words "so they can leverage its full frontier-level cybersecurity defense capabilities". Google's post describes safeguards covering misuse prevention, prompt injection resilience, misalignment monitoring and hardened sandboxed testing before broader availability.

The sequencing mirrors what Anthropic argued the same week in its report on an open Chinese model. We covered that report in our piece on GLM-5.3's cyber capabilities. Its recommendation was that cyber defenders should get access to frontier models comparable to the ones attackers can reach. Google is, in effect, doing that, with a vetted list as the control.

The second gate is political. VentureBeat reports that developers and enterprises must wait for what it describes as "the U.S. government's voluntary pre-release model access process". That process follows the White House's voluntary accord, which we explained in our guide to what the accord commits companies to. Argon is among the first launches where the commitment visibly shapes a release schedule.

How it compares on price

The price is aggressive on paper. VentureBeat reports GPT-6 Astra is priced at $10 and $50 per million input and output tokens, so Argon's introductory rate is about a fifth of that, and its standard rate of $4 and $20 would still be well under half. It matches Claude Opus 5.5's base rates, according to the same report. Output length also changes the arithmetic. A model allowed to write 1 million tokens in one answer can run up a large bill in a single call, so a cheap per-token rate is not the same as a cheap task.

What you can and cannot do with it

As of 1 October 2026, DataCamp reports there is no published API model ID for Argon, and the model does not appear on OpenRouter, Vertex AI or other major platforms. So the price list is a price list for a model most developers cannot call.

Google also says thousands of its own staff already use it, including for optimising quantum algorithms, freeing more than 300 terabytes of memory across data centres, and migrating C and C++ code to Rust. Those are internal results, unverifiable from outside, and worth noting as a sign of where Google thinks the value is: cost-saving engineering work at scale, not chat.

The UK reader's position

For a UK team, nothing changes this week. Fairwind is a trusted-defender programme and the pre-release access process named by VentureBeat is American. We have not seen a statement on UK availability, and we will not guess at one. If you buy Google Cloud in sterling, expect to convert the dollar prices and to wait.

Our take

Argon is a serious release and an unfinished one. Google took the lead on a majority of its own chosen benchmarks, then limited access to the one group where a mistake is most expensive. That is a defensible way to ship a model with strong cyber capability, and it is also a way to postpone independent testing.

What we would watch is the first third-party evaluation, the published model ID, and what the post-introductory price turns out to be. Until all three exist, the right comparison is not Argon against GPT-6 Astra or Claude Opus 5.5. It is Google's claims against the evidence that has not yet arrived.