Google unveils Gemini 4 Argon, but the debut happens behind closed doors

On September 30, 2026, Google announced Gemini 4 Argon, the first model of the Gemini 4 generation, in a company blog post signed by Koray Kavukcuoglu, senior vice president at Google DeepMind and head of AI architecture. The initial rollout follows a strategy that has become common among AI labs: immediate access is limited to cyber defenders selected through the company's Fairwind Program. For developers, enterprise customers, API users and Google AI Ultra subscribers, the release is described as coming 'as soon as possible', with no date given, and the available sources include no details or timeline for availability in the European Union. According to VentureBeat, the Fairwind Program provides the model without its cyber guardrails to trusted defenders and to Google's internal teams. Wiz is already using it in its Scan for Good initiative, which focuses on protecting critical infrastructure.

On performance, Google says the output limit rises to one million tokens, up from 64,000, and claims frontier results in software engineering, enterprise knowledge work — legal and financial — and cyber defense. By a tally VentureBeat published across 18 cited benchmarks, Argon ranks first in 12, ahead of competitors such as Claude Opus 5.5 and GPT-6 Astra on tests like DeepSWE v1.1 (77.9%) and AutomationBench (51.3%). The same outlet notes, however, that Argon does not lead on every metric: GPT-6 Astra stays on top on FrontierSWE v2 and Terminal-Bench Science 0.1, while Claude Opus 5.5 beats Argon on Terminal-bench 4.0. As TechCrunch points out, these are self-reported scores from the vendor and still await full independent verification.

Among the internal use cases Google cites is the migration of more than 800,000 lines of code from C/C++ to Rust for Zircon, the Fuchsia kernel. Separately, the company says that in a demonstration run with Wiz the model found a critical vulnerability in healthcare software that earlier frontier models had missed; it was not described in detail. On pricing, Google has set an introductory rate of $2 per million input tokens and $10 per million output tokens. According to VentureBeat, that is a fifth of GPT-6 Astra's price ($10 and $50) and half of Claude Opus 5.5's ($4 and $20). Once the introductory period ends, and its length is unknown, the price rises to $4 for input and $20 for output, exactly Opus 5.5's list price. Cached input is discounted by 95%.

Reserving the most capable models for defensive use before a mass rollout is becoming a recurring choice in the industry. Still unknown: when access opens to everyone else, how long the introductory price will last, and the criteria Google uses to pick the defenders who receive the model without cyber guardrails. — Olya