Google has begun rolling out Gemini 4 Argon, which it calls its next frontier model, to a set of vetted cyber defenders through its Fairwind Program, with developers, enterprises and consumers to follow. The announcement, signed by DeepMind's chief AI architect Koray Kavukcuoglu, is packed with the numbers that now define a frontier launch: a new state of the art on a long-horizon software-engineering benchmark, leading scores across finance and legal evaluations, an "industry-leading" output ceiling, and an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%. Read plainly, it is a strong release. Read carefully, it is also a document about who gets frontier capability first, and why.
The output limit is the real headline
Argon's output token limit rises to 1 million tokens, up from 64K. Input windows have grown for years; the output side has stayed comparatively cramped. That asymmetry is what forces agents to loop: write a little, act, observe, write a little more. Every loop is a place where a long task can derail or drift. A 1M-token single trajectory does not by itself make a model smarter, but it changes the shape of the problems you can hand over in one pass: an entire codebase migration, a full research write-up, a document workflow spanning dozens of inputs. The cost implication is equally concrete. At $10 per million output tokens, a maximum-length single output costs about $10 before any input, which turns "how long should the model think?" into a budget line rather than a setting.
The internal results are impressive, and unverifiable
Google cites a quantum-algorithm subroutine that beat a published baseline by 40% "in a matter of minutes"; a fleet-wide agent effort that freed more than 300 TiB of data-center memory, with an estimated 500 TiB to 1 PiB in total savings; and C/C++-to-Rust migrations scaling up to 800,000-plus lines for the Fuchsia Zircon kernel. The most detailed is libgav1, where Argon agents replaced 32,000 lines of SIMD code with safe Rust built to auto-vectorize, producing a memory-safe video decoder that runs 2.7x faster than the prior Rust port with identical output.
These are exactly the kind of claims that are hard to check. They were produced on Google's systems, against Google's baselines, by Google's teams, and the underlying code and telemetry are not public. The quantum figure compares against "the published baseline," not against a competing model under the same conditions. The memory savings are "estimated." None of this makes the claims wrong — several look like the most substantive evidence in the post — but they are vendor numbers, and the history of frontier-model announcements is a history of vendor numbers that shrank when outsiders tried them.
The public benchmarks are more comparable, and worth reading closely
Argon sets a new state of the art on DeepSWE v1.1 at 77.9%, which measures real long-horizon software-engineering tasks, and leads the Vals Index, a composite that weights finance, coding, legal and tax work by each sector's contribution to U.S. GDP, alongside domain evaluations such as Vals Finance Agent v2 and Harvey's legal-agent benchmark. On LVBench, for long-video understanding, it scores 91.7%.
The number that deserves more attention is easy to miss. On AutomationBench, Zapier's benchmark for end-to-end execution across core business functions, Argon "ranks #1 with a score of 51.3%." A leaderboard-topping score of barely half is a healthy reminder that even the best frontier model still fails about half the time on realistic multi-step office work. That single figure may say more about where agentic AI actually stands than any of the impressive internal wins.
Capability and policy now ship as one package
Argon goes first to cyber defenders, under a phased release that Google explicitly ties to the U.S. government's voluntary pre-release model-access process. For trusted defenders and Google's own teams, it is released without cyber guardrails so they can use the model's full reach. Google says Argon can autonomously find, validate and patch critical vulnerabilities; it cites Wiz's Scan for Good program, which used the model to surface a critical vulnerability in healthcare software used by hospitals worldwide that earlier frontier models missed, and a 68% top score on CWE-bench v1.
The structure here is the story. A frontier model reaching government-linked defenders first, inside a pre-release review regime, is becoming the default path for the largest labs rather than an exception. The same announcement that promises a security tool also normalizes a release gate, and ships a model with its cyber guardrails removed for some users while they stay on for everyone else. That is a defensible trade for defense work, but it is a governance choice, made by a vendor, and it deserves to be named as one.
Where the tension shows
Google lists four hardening areas: refusing harmful cyber and CBRN requests while preserving dual-use research; stronger resistance to indirect prompt injection, which it claims leads on Gray Swan's IPI benchmark; monitoring Argon's chain-of-thought and actions for "misalignment," stopping execution when needed; and sealing sandboxed environments before high-risk training or evaluation. The company also urges the rest of the industry to "preserve reasoning transparency" so model thoughts stay inspectable. It is a thoughtful list, and it is also the list a company publishes when it is selling a model whose most valuable use — autonomous vulnerability discovery and patching — sits one step from the misuse it is promising to prevent.
Our take
Gemini 4 Argon looks like a genuine frontier step, and the 1M-token output window is the change most likely to matter, because it relaxes the loop length that has quietly capped what agents can do. But the release is as notable for its shape as for its scores. The compelling evidence is largely first-party and unreproducible; the most telling public number is a 51.3% top score; and the model arrives gated by a pre-release review process and split, by trust level, into guarded and unguarded versions. Frontier capability and frontier policy are now bundled at launch. Treat the benchmarks as an opening bid, and watch for the replication — and for what it means that the first customers for the most capable cyber model are defense teams on either side of a government process the vendor is also helping to shape.
Sources
- Google, "Gemini 4 Argon: our next era of frontier intelligence," Google blog (Sep 30, 2026) — https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/