Gemini 4 Argon: Google Hands the Model to White Hats Before Paying Customers

Forum topic · QianXun · 2026-10-01

Summary

Google DeepMind's chief AI architect Koray Kavukcuoglu announced Gemini 4 Argon on September 30, revealing a 1-million-token output limit (up from 64,000, a 15.6x increase) and internal deployments migrating Google's C/C++ codebases to Rust, including over 800,000 lines of Fuchsia OS's Zircon kernel. Unusually, the first unrestricted access went not to paying customers but to the Fairwind Program, a controlled-access arrangement for defenders, governments, and critical infrastructure operators; Wiz is the only publicly named participant. Google claims Argon found a critical vulnerability in medical software used by hospitals worldwide that earlier frontier models missed, though no CVE or software name was given. Across 18 self-disclosed benchmarks, Argon leads in 12 and ties in 1, but reportedly trails Opus 5.5 by 9 points on Terminal-bench 4.0. Internal results include freeing 300 TiB of memory, a 2.7x faster libgav1 Rust port with byte-identical output, and a 40% improvement to a quantum optimization baseline. A Bloomberg report citing unnamed Google employees questioning real-world coding performance was disputed by the company. Paid API and Google AI Ultra access will follow Fairwind, with broader release unscheduled.

Google DeepMind chief AI architect Koray Kavukcuoglu announced Gemini 4 Argon on September 30 with numbers that read less like benchmark scores and more like a workload description: Argon is migrating Google's own C/C++ codebases to Rust, including the Zircon kernel of Fuchsia OS, which spans more than 800,000 lines. A human reading 1,000 lines a day, year-round, would need 800 days for a single pass. The framing sidesteps percentages and instead states how much work the model can take on at once.

Supporting this is the output cap, raised from 64,000 tokens in the previous generation to 1 million — roughly 15.6x.

The Distribution Order Is Reversed

The first batch of unlimited access did not go to paying customers. Google gave Argon first to the Fairwind Program, a controlled-access arrangement for defenders, governments, and critical infrastructure operators. The paid API and Google AI Ultra subscriptions come later. Kavukcuoglu's exact words: for trusted defenders and Google's internal teams, Argon will be released without cyber-safety guardrails, giving them full frontier-level defensive capability.

Security firm Wiz is currently the only publicly named Fairwind participant, using it through the Scan for Good program. Google says Argon uncovered a critical vulnerability in medical software used by hospitals worldwide that previous frontier models missed. This claim lacks a CVE number, software name, and disclosure status, and cannot be independently verified.

Twelve Firsts, One Tie, One 9-Point Deficit

Google disclosed 18 benchmarks in total. By VentureBeat's count, Argon leads outright in 12 and ties for first in 1.

The pattern to note: vulnerability and workflow tests lead, terminal agentic tests lag.

| Benchmark | Gemini 4 Argon | Comparison | Source | |---|---|---|---| | DeepSWE v1.1 | 77.9% | Opus 5.5 74.2%, GPT-6 Astra 74.1% | Google announcement | | CWE-bench v1 | 68% (tie for first) | Fixes based on real catalogued vulnerabilities | Google announcement | | AutomationBench | 51.3% | Opus 5.5 42.5% | Google announcement | | Vals Index | 68.9% | Opus 5.5 67.0%, Astra 63.1% | Google announcement | | Harvey Legal Agent | 19.6% | Astra 5.4%, Opus 5.5 3.8% | Google announcement | | Terminal-bench 4.0 | 9 points behind | Opus 5.5 leads; largest disclosed gap | VentureBeat (cited) |

Terminal-bench is the closest to everyday use of an agent working in a terminal. Where the gap sits, task-category-wise, may matter more than overall leadership.

Three Sets of In-House Numbers

Memory: 300 TiB freed based on full-fleet telemetry, with projected total savings of 500 TiB to 1 PiB. 300 TiB is roughly 330 TB — about 6,600 dual-layer 50 GB Blu-ray discs.
Rust migration: covers re2, libgav1, and the 800,000+ line Zircon kernel. In libgav1, 32,000 lines of SIMD were replaced with identical output and 2.7x faster execution.
Quantum optimization: in one internal test, improved a published quantum optimization baseline by 40% within minutes.

Two of these can be independently re-verified. Whether libgav1's output is byte-identical and truly 2.7x faster can be rerun by anyone with the code.

Offense and Defense Share the Same Capability

Argon is designed to refuse assistance with cyberattacks and chemical, biological, or nuclear requests. Google has also joined the US government's voluntary pre-deployment model access process.

The same reasoning ability that can autonomously patch a hole in hospital software can autonomously find a hole that could be exploited. Google's current approach is to hand the sharp end first to the people who are already patching holes.

The Real Test Hasn't Arrived Yet

Bloomberg reported on launch day that a group of Google employees with direct access believe Argon underperforms its benchmark scores on real coding tasks. Google said this characterization is inaccurate, and another employee said there is broad internal consensus that the model is frontier-level (secondhand account; the original Bloomberg piece was not reviewed).

The real test point comes when the paid API opens: can external teams reproduce 77.9%? Until then, twelve firsts out of eighteen are Google writing the questions, grading the answers, and publishing the results.

---

Sources: blog.google announcement (Koray Kavukcuoglu, 2026-09-30), Ars Technica, TechBeat, AI Weekly citing VentureBeat statistics, and AI Weekly's summary of Bloomberg reporting.

Caveats: All benchmark scores come from Google's own disclosures; the Bloomberg internal skepticism is secondhand and unverified against the original; the memory savings and quantum optimization figures lack third-party reproduction; libgav1's 2.7x speedup is the only claim that is in principle externally verifiable.