Google DeepMind Unveils Gemini 4 Argon, a Frontier Model for Cyber Defenders First

Google announced Gemini 4 Argon on September 30, 2026, introducing its first new flagship Gemini generation since Gemini 3 and framing the release as a step change in frontier capability. The company is not handing the model to everyone at once. Instead, Argon is rolling out first to a set of trusted cyber defenders through Google's Fairwind Program, an arrangement that puts security teams at the front of the queue while developers, enterprises and consumers wait.

The launch lands in a crowded autumn for frontier AI. Gemini 3, released toward the end of 2025, first put Google back in the frontier conversation, but Anthropic and OpenAI led much of 2026. Argon arrived one day after OpenAI's DevDay conference, where OpenAI launched its Dots AI agent and the GPT-6.1 Sol model, and also said it would not release a planned GPT-6.1 Astra model because of safety worries. Google's choice to lead with defenders is both a technical claim and a positioning move.

The headline numbers are aggressive. Argon expands the output token limit to 1 million tokens, up from the 64K tokens that earlier Gemini models topped out at. At launch it costs $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price. A Google footnote states that pricing rises to $4 and $20 per million tokens after the introductory period, which is exactly what Anthropic charges for Claude Opus 5.5.

Access is narrow by design. SiliconANGLE reported on September 30 that outside Google's own teams, only members of the Fairwind Program can use Argon for now. That program opened on September 3 with the smaller Gemini 3.8 Flash Cyber model and has signed up more than 650 organizations, including CrowdStrike Holdings Inc. and Palo Alto Networks Inc. Koray Kavukcuoglu, Google's chief AI architect and DeepMind senior vice president who became head of DeepMind in August 2026, wrote that releasing capabilities at this level requires a phased approach, and Google is taking part in the U.S. government's voluntary pre-release model access process.

Key Facts

Google describes Argon as delivering frontier performance in complex workflows across real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. The Verge reported on September 30 that the model is already powering Google's internal workflows, including large-scale codebase migrations, and that Google shared benchmark charts showing Gemini 4 outperforming competing models from OpenAI and Anthropic.

The benchmark picture is strong but not a clean sweep. SiliconANGLE reported on September 30 that Argon scored 77.9% on DeepSWE v1.1 for long-horizon software engineering tasks, against 74.2% for Anthropic's Claude Opus 5.5, which itself sat a tenth of a point ahead of OpenAI's GPT-6 Astra. On AutomationBench, which measures end-to-end business work, Argon led by 51.3% to 42.5%. Of 18 benchmarks on Google's charts, VentureBeat counted 12 that Argon led outright, while Anthropic's Opus 5.5 still held Terminal-Bench 4.0 and OpenAI's Astra kept FrontierSWE v2. On CWE-bench v1, a remediation test, Argon tied GPT-6 Astra at 68%.

Inside Google, the model has already been put to work on unglamorous but expensive problems. Google wrote on September 30 that Argon helped quantum researchers beat a published baseline by 40% in minutes, that Argon agents freed over 300 TiB of data-center memory against an estimated 500 TiB to 1 PiB in total, and that agents migrated C and C++ code to Rust, scaling from tens of thousands of lines in re2 and libgav1 up to more than 800,000 lines for the Fuchsia Zircon kernel. For libgav1, agents replaced 32,000 lines of SIMD code and produced a memory-safe decoder that runs 2.7 times faster than the prior Rust port with identical video output.

Security work is where the release is most pointed. SiliconANGLE reported on September 30 that Google-owned Wiz used Argon in its free Scan for Good program and found a critical flaw exposing personal information in health care software used by hospitals worldwide, a flaw that earlier frontier models had missed. Fairwind members and Google internal teams receive a version with the cyber guardrails removed, which is how defenders get the model's full offensive and defensive range rather than a softened build.

The Verge reported on September 30 that Google will strengthen critical frontier safeguards before a broader rollout, including defenses against misuse and prompt-injection attacks and monitoring for misalignment. That sequencing, defenders first, developers and consumers later, is the core of the launch story.

Analysis

What this really means is that Google is treating frontier capability as something to be staged rather than shipped. A 1 million token output limit and benchmark wins are the easy part to advertise; deciding who gets the model first, and under what safeguards, is the harder and more consequential choice. A model that is strong in cyber defense is also a model that can be pointed at targets, and Google has chosen to put it in defenders' hands before anyone else's.

CNBC reported on October 2 that analysts are bullish while noting the real test will come when businesses can deploy the model widely. Tim Law, director of research for AI at IDC, said that Gemini 4 Argon shows advanced reasoning on critical tasks, particularly legal reasoning, finance and other aspects of enterprise knowledge work, including long-running tasks. Lian Jye Su, chief analyst at Omdia, said Google was late to the cybersecurity and coding game but that Gemini 4 took the company to the frontier, particularly in cybersecurity. Those are qualified endorsements rather than coronations.

The qualification matters because the competitive scoreboard is not unambiguously Google's. CNBC reported on October 2 that the Artificial Analysis Intelligence Index, a composite benchmark, places Gemini 4 behind only Claude Opus 5.5 and Claude Sonnet 5.5 on its leaderboard. Nick Patience, AI lead at the Futurum Group, told CNBC that Argon makes Google competitive again but does not make it the leader. That is the honest read of a launch where Google won 12 of 18 published charts while two of the most visible benchmarks stayed with rivals.

The bigger picture here is that the frontier race has split into two contests. One is raw benchmark leadership, where margins are thin and the leaderboard changes monthly. The other is deployment, safety process and trust, where Google is trying to turn a government pre-release review and a 650 organization defender program into an advantage. Pricing reinforces the point: at $2 and $10 per million tokens, Argon undercuts Anthropic's Opus 5.5 by half during the introductory window, then matches it at $4 and $20. Tulsee Doshi, Google's Gemini model product lead, told CNBC that starting the rollout in phases gives the company more confidence but also lets it put a model strong in cyber defense in the hands of defenders as soon as possible.

Why It Matters

For security teams, Argon arriving first is a meaningful shift in how frontier AI reaches practitioners. The Fairwind Program began on September 3 with Gemini 3.8 Flash Cyber and has grown past 650 organizations, including CrowdStrike and Palo Alto Networks. Those members get a build with cyber guardrails removed, and the Wiz Scan for Good result, a critical flaw in hospital software that earlier frontier models missed, is the kind of concrete outcome that justifies the sequencing. If defenders can find and fix flaws earlier, the asymmetry between attackers and defenders narrows.

For enterprises, the launch is a promise rather than a delivery. Argon is not yet generally available to developers, enterprises or consumers, and Google has given no date for paying developers and Google AI Ultra subscribers, who are next in line. The legal, finance and long-running agent workloads that Google and IDC highlight will only be tested at scale once those customers can run the model in production. Tim Law of IDC said the final proof will be in enterprise production environments once the model is fully released.

For Google itself, Argon is a test of whether a staged release can rebuild frontier credibility. Gemini 3 opened the door in late 2025, and Anthropic and OpenAI walked through much of 2026. Argon, with a 1 million token output ceiling, 12 outright wins out of 18 published benchmarks, and a defender first rollout, is Google's attempt to set the terms again. The company's own safeguards commitments, defenses against misuse and prompt-injection attacks and monitoring for misalignment, will be judged alongside the benchmark charts.

Next Up

The immediate next step is expansion. Google says it will gradually widen access to developers, enterprises and consumers, and it is engaged in the U.S. government's voluntary process for pre-release model access. No date has been given for the broader release or for when paying developers and Google AI Ultra subscribers get access, so the practical question for most builders is timing rather than capability.

Two other clocks are running. The first is price: introductory rates of $2 per million input tokens and $10 per million output tokens give way to $4 and $20, with cached input still 95% off the input price, so teams that adopt early hold a cost advantage that will not last. The second is competitive: OpenAI's DevDay produced the Dots agent and GPT-6.1 Sol, a planned GPT-6.1 Astra was withheld over safety worries, and Anthropic still holds the top two slots on the Artificial Analysis Intelligence Index. Argon's defenders first strategy will be tested as soon as the model reaches everyone else.