← Back to feed

Google released Gemini 4 Argon on September 30, its first genuine frontier model since the Gemini 3.x generation, and immediately restricted access to a narrow cohort of trusted cybersecurity researchers through the Fairwind Program. The rollout to standard API customers and Google AI Ultra subscribers is described as coming “as soon as possible” pending further evaluation.

The model is built around long-horizon workflows. Its headline technical addition is a 1M output token limit, up from 64K on predecessor models, which allows Argon to sustain deep reasoning chains, large-scale code migrations, and multi-document analysis without truncating mid-trajectory.

Benchmark Position

On DeepSWE v1.1, which measures performance on real-world long-horizon software engineering tasks, Argon posts 77.9% — the highest published score on that benchmark. GPT-6.1 Sol, announced the day before, reports 75.2% on the same benchmark at high reasoning effort.

On AutomationBench, Zapier’s end-to-end execution benchmark across core business functions, Argon ranks first with 51.3%. GPT-6.1 Sol scores 31.7% at medium reasoning effort on the same benchmark.

On LVBench, which measures long video understanding, Argon claims 91.7%, reported as state of the art.

On CWE-bench v1, evaluating vulnerability remediation, Argon ties for first place at 68%, sharing that position with GPT-6 Astra and Grok 4.7.

Google also cites leading positions on the Vals Index (economic impact across finance, coding, legal, and tax), Vals Finance Agent v2, and Harvey’s Legal Agent Benchmark for legal research and drafting, publishing a comparison table with numeric scores against GPT-6 Astra and Claude Opus 5.5 across those benchmarks.

Fairwind Access and Phased Release

The Fairwind Program is Google’s controlled early-access channel for cybersecurity work. The same program governed early access to Gemini 3.8 Flash Cyber, which gained attention earlier in September for its role in Wiz’s Scan for Good vulnerability discovery initiative.

Argon extends that framing. Google describes the model as capable of finding, validating, and patching software vulnerabilities, and cites a collaboration with Wiz’s Scan for Good programme — including a specific claim about discovering a critical exposure in healthcare software that earlier frontier models missed. That example is attributed to Wiz and Google and has not been independently reproduced from public technical evidence.

The U.S. government’s voluntary pre-release model access process is also cited. Google says it is “actively engaged” in that process as it expands access.

Pricing

Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the standard input rate. After the introductory period, pricing rises to $4 per million input and $20 per million output.

At introductory rates, Argon sits at the same input token price as GPT-6.1 Sol ($2/M), though Sol’s output price is also $10/M — making the two models equivalent at launch pricing per token. Argon’s post-introductory price doubles both figures.

What Changes

The previous Gemini 4 signal, from a September 23 public statement by SVP Koray Kavukcuoglu, indicated only that the model was in post-training and would ship “much earlier than the end of the year” — without a specific quarter target. Argon is the delivery on that commitment, arriving within a week of that statement, and with specific benchmark claims attached.

The 1M output token limit is the most structurally significant specification. Most frontier models cap output at 32K to 64K tokens. Extending that to 1M enables single-trajectory runs that previously required multi-session chaining — particularly relevant for the long-horizon coding and enterprise knowledge workflows Argon is positioned around.

SWE-bench Verified scores and tau-bench data are not yet available. The DeepSWE, AutomationBench, and LVBench figures are from Google’s own launch materials and have not yet been independently reproduced on standardised harnesses.