Google drops Gemini 4 Argon, its most capable model yet

Google isn't playing catch-up anymore. On September 30, 2026 (UTC), the company unveiled Gemini 4 Argon — its most capable model to date, built for deep, long-horizon reasoning across coding, enterprise knowledge work, and cybersecurity defense. The model ships with an industry-leading 1 million token output limit, a price point half of Claude Opus 5.5, and a phased rollout that starts with trusted cyber defenders rather than the general public.

What Argon brings to the table

The headline number is the output window: 1 million tokens, up from 64K in previous Gemini models. That's a 15x jump, and it matters because frontier models increasingly need room to think — generating hundreds of thousands of tokens in a single trajectory to solve multi-step problems without losing the thread.

On benchmarks, Argon delivers across the board:

Sources: Google DeepMind official blog, 9to5Google. DeepSWE figures from Google announcement; CWE-bench tie includes Grok 4.7.

The DeepSWE gap is worth unpacking. Argon's 77.9% beats Claude Opus 5.5 by 3.7 percentage points and GPT-6 Astra by 3.8 points. That may sound modest, but at the frontier level — where models are already solving 70%+ of real-world engineering tasks — each point represents increasingly hard problems that separate "useful copilot" from "autonomous engineer." Google is claiming the crown in the benchmark that most directly maps to developer productivity.

Pricing is equally aggressive. Argon launches at $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. That's exactly half of Claude Opus 5.5's $4/$20 pricing. For a model that beats Opus on the flagship coding benchmark, undercutting it by 50% is a deliberate market-share play — Google is signaling that frontier intelligence doesn't have to come at a frontier premium.

Already running inside Google

Google didn't just build Argon in a lab and ship it. The model is already deployed internally at scale, and the use cases are concrete:

• Memory optimization: Argon agents analyzed fleet-wide profiling telemetry and autonomously applied memory optimizations across Google's data centers, freeing over 300 TiB of memory so far, with an estimated 500 TiB to 1 PiB in total savings. No new hardware purchased — just smarter code.
• Quantum algorithm optimization: Argon helped quantum researchers optimize spacetime resources (qubits × gates) of bottleneck subroutines. In one example, it beat the published baseline by 40% in minutes.
• Large-scale Rust migration: Argon agents are migrating C/C++ codebases to Rust across Google — from tens of thousands of lines in core libraries like re2 and libgav1 up to 800,000+ lines for the Fuchsia OS Zircon kernel. For libgav1, Argon replaced 32K lines of SIMD code with safe Rust that the compiler auto-vectorizes, producing a memory-safe decoder that runs 2.7x faster than the existing Rust port with identical video output.

These aren't demo cases. They're production workloads at one of the largest engineering organizations on Earth. The 800K-line Zircon kernel migration, in particular, is a scale of automated code rewriting that would have been unthinkable two years ago — and it's undergoing rigorous auditing before production rollout, which is the right call.

Cybersecurity as a first-class capability

Argon's cybersecurity positioning is the most strategically interesting part of this launch. Google trained the model specifically for defensive cybersecurity — it can autonomously find, validate, and patch critical vulnerabilities. For trusted defenders and Google's own internal teams, Argon is being released without cyber guardrails, unlocking its full frontier-level security capabilities.

The early results are striking. Cloud security firm Wiz is already using Argon through its "Scan for Good" program, which protects critical public infrastructure for free. In an early demonstration, Argon uncovered a critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide — a severe risk that previous frontier models had missed.

On CWE-bench v1, which measures vulnerability remediation, Argon ties for first place at 68% alongside GPT-6 Astra and Grok 4.7. On Google's internal vulnerability benchmark, it found exposures across 20 programming languages. On Wiz's black-box penetration testing benchmark — analyzing live web systems without source code — it outperforms 3.8 Flash Cyber across attack surface discovery, vulnerability identification, and proof-of-concept validation.

The strategic implication is clear: Google is positioning Argon as the model of choice for cybersecurity defenders, not just another general-purpose LLM. By gating the full cyber capabilities to trusted partners through the Fairwind Program, Google is also addressing the dual-use problem head-on — the same capabilities that find and patch vulnerabilities can be weaponized to exploit them.

Why it matters

Gemini 4 Argon marks Google's return to the frontier conversation on its own terms. For the past year, the narrative has been set by OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5, with Google focused on faster, cheaper Flash models. Argon changes that — it's a direct shot at the high-end market, and it arrives with a price/performance ratio that neither competitor can match on paper.

The 1 million token output window is more than a spec bump. It represents a philosophical shift in how Google thinks about model capability: give the model room to reason deeply, and it can solve problems that require sustained, multi-step thinking rather than quick pattern matching. This aligns with the industry's broader move toward agentic systems — models that don't just answer questions but execute complex workflows over extended periods.

There's a critical tension worth naming. Google is releasing Argon's full cybersecurity capabilities to trusted defenders without guardrails, while simultaneously strengthening safeguards against misuse, prompt injection, and misalignment for the general release. This dual-track approach — maximum capability for vetted partners, maximum safety for everyone else — is becoming the industry standard, but it raises questions about who gets access to frontier capabilities and who decides. The Fairwind Program's vetting criteria and governance structure deserve scrutiny.

The phased rollout also tells us something about Google's confidence. By starting with cyber defenders and engaging the U.S. government's pre-release review process before broader availability, Google is being more cautious than OpenAI was with GPT-6 Astra's DevDay launch. Whether this is genuine safety prudence or a response to the recent wave of agentic security incidents — the Hugging Face breach, the Australian Medicare intrusion, the DIVD AI agent attack — is a fair question. The timing, one day after Pichai signed the White House voluntary safety accord, is not a coincidence.

Watch three things next: whether developers report real-world coding gains matching the DeepSWE numbers; whether the half-price pricing forces Anthropic and OpenAI to cut Opus/Astra prices; and how quickly Google expands beyond the Fairwind Program to general availability. The frontier model race just got a serious third contender, and the pricing pressure is about to get real.