Gemini 4 Argon ships first to cyber defenders, not the public API

Google announced Gemini 4 Argon on September 30, 2026, and almost no one outside a narrow test group can use it. This is a gated cyber-defense rollout, not a public API, enterprise launch, or consumer release.

The Verge reported that Google is limiting first access to a “set of trusted cyber defenders.” Chief AI architect and Google DeepMind SVP Koray Kavukcuoglu said the company is “actively engaged in the U.S. government’s voluntary process for pre-release model access while we gradually expand access.” Ars Technica put the same phase in program terms: partners in Google’s Fairwind Program can use Argon for cybersecurity defense. There is no public date for everyone else. After those tests, general availability is supposed to start with paid API users and Google AI Ultra subscribers, and only later reach enterprise and consumer customers. Until then, agent builders cannot call Argon.

Kavukcuoglu described the model as delivering “frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense.” CEO Sundar Pichai, posting the same day, said it shows that kind of performance in complex workflows, cyber defense, and software engineering, and that teams at Google are already using it extensively. That is internal access. It is not a developer key.

Intro prices and a 1 million token output cap

The model is still in limited testing, but Google has published introductory API rates. For a limited time, Argon is $2 per million input tokens and $10 per million output tokens, and cached input tokens are discounted 95 percent. Neither report gives the rate after that window, so the post-intro price is unknown.

Google also said the output limit rises to 1 million tokens, from 64,000 tokens in previous Gemini models. The company says the larger cap lets users finish more demanding tasks in a single step. For long-horizon agent work, that output budget is the concrete spec on the table — and the endpoint is not.

Benchmarks and internal use, as Google tells them

On the DeepSWE v1.1 software-engineering benchmark, Google says Argon scores 77.9 percent, higher than GPT-6 Astra, Fable 5.1, and Opus 5.5. It also pointed to an industry-leading result on the economic-analysis Vals Index. The published accounts do not give that Vals score. The Verge noted a chart in which Gemini 4 fares better than models from OpenAI and Anthropic, after apparent benchmark figures had already circulated on X earlier in the week. Those figures are Google’s claims, not independent measurements.

Inside the company, Ars Technica reported that Argon used “fleet-wide telemetry data” to help save 300 TiB of memory across Google’s data centers. Argon agents have also been migrating C and C++ codebases to Rust, including thousands of lines in the re2 and libgav1 libraries and more than 800,000 lines in the Fuchsia OS Zircon kernel. The Verge separately noted large-scale codebase migrations among the internal workflows Kavukcuoglu said the model is already powering.

A cyber-first gate, with monitors

Cyber defense is the point of this first phase. Google says Wiz is already using Argon and used it to uncover a critical vulnerability that could expose personal information in a system used at hospitals around the world. Google claims other frontier models missed the flaw, and it did not provide specifics. Read that as Google’s account.

Google says it designed Argon with systems that monitor the model’s chain-of-thought and can stop it if the model steps out of bounds, and it tied that design to the case for reasoning transparency. The Verge reported that Google will strengthen “critical frontier safeguards” before a broader rollout, including defenses against misuse and prompt injection and monitoring for misalignment. The Verge framed the limited launch as a way for Google to make sure the model is not misaligned.

The announcement landed a day after OpenAI’s DevDay, where OpenAI launched its Dots agent and the GPT-6.1 Sol model. The same week, OpenAI said it would not release a planned GPT-6.1 Astra model over safety worries. Argon also follows Kavukcuoglu’s appointment as head of DeepMind in August. Ars noted the longer path: Google had promised Gemini 3.5 Pro in June, then spent the summer on smaller Flash models.

If you buy API capacity or design long-running agents, read the Verge and Ars accounts before you put Argon on a roadmap. Budget against the stated intro prices and the 1 million token output cap, not against a public endpoint. Watch for the paid-API and Google AI Ultra phase, and keep the DeepSWE score and the Wiz finding labeled as Google’s claims until someone outside the company can check them.

Intro prices and a 1 million token output cap

Benchmarks and internal use, as Google tells them

A cyber-first gate, with monitors

Sources