Google announces Gemini 4 Argon AI model, limits early access to cybersecurity defenders

Google announced a new artificial intelligence model, Gemini 4 Argon, on 30 September 2026. Ars Technica and The Verge report that the model is not yet generally available and that access is initially limited. Anadolu Agency's headline says Google has begun rolling the model out to cybersecurity partners.

The Verge and Anadolu Agency report that the first users are cybersecurity partners or defenders. The Verge quoted Google as describing them as a "set of trusted cyber defenders". Google also released benchmark results comparing the model with competing models, including models from OpenAI and Anthropic.

Pricing and features

According to Ars Technica, Google announced limited-time API pricing of $2 per million input tokens and $10 per million output tokens. Cached input tokens get a 95 percent discount. Ars Technica also reports that the model will support an output limit of 1 million tokens, up from 64,000 tokens in previous Gemini models.

Internal use

According to Ars Technica, Argon agents have been working to migrate C/C++ code to Rust across Google. The effort covers thousands of lines in the re2 and libgav1 libraries and more than 800,000 lines in the Fuchsia OS Zircon kernel. Ars Technica also reports that Argon reportedly used "fleet-wide telemetry data" to help Google save 300 TiB of memory across its data centres.

Context

According to The Verge, the announcement came one day after OpenAI's DevDay conference. There, OpenAI launched its Dots AI agent and its GPT-6.1 Sol model. The Verge also reports that OpenAI said this week it would not release a planned GPT-6.1 Astra model because of safety concerns. Earlier in the week, what appeared to be leaked benchmarks for Google's model circulated on X, The Verge reports. Ars Technica reports that Google promised Gemini 3.5 Pro in June but spent the summer releasing smaller Flash models.

What each side says

Google says the model offers industry-leading performance in coding, knowledge work and cybersecurity. According to Ars Technica, Google says it scores 77.9 percent on the DeepSWE v1.1 software engineering benchmark. Google says this is higher than GPT-6 Astra, Fable 5.1 and Opus 5.5. Google says the 1 million-token limit lets users complete more demanding tasks in a single step.

The Verge quoted Koray Kavukcuoglu, whom it identifies as Google's chief AI architect and a Google DeepMind SVP, as saying the model shows "frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense". He said Google is "actively engaged in the U.S. government’s voluntary process for pre-release model access while we gradually expand access." According to The Verge, Google says it will strengthen safeguards before a broader rollout. These include defences against misuse and prompt injection attacks and monitoring for misalignment.

The reports do not include responses from OpenAI, Anthropic or other competitors named in the comparisons.

What is not yet known

None of the reports describe independent verification of Google's benchmark results or internal-use figures. The reports do not give a timeline for general availability. They do not identify the selected cyber defenders or explain how they were chosen.

Analysis

All three outlets ultimately rely on Google's own announcement. Its performance and efficiency claims therefore remain company statements until they are independently tested. The limited first release is notable. It comes in the same week that OpenAI said it would hold back a planned model over safety concerns. This may suggest that major developers are giving more weight to staged access for capable models, although the reports do not establish a link between the two decisions. Possible next steps include outside benchmark testing, details on the U.S. government's pre-release process, and a timeline for wider access. Each could confirm or complicate Google's account.

Produced by the Vocemundi newsroom with AI assistance.