Google's Gemini 4 Argon is here — but only 'trusted cyber defenders' get access first

Google has officially unveiled Gemini 4 Argon, its next frontier model — but most developers won't get to touch it yet. The company announced on September 30 that initial access will be limited to a "set of trusted cyber defenders," with a broader rollout dependent on completing "critical frontier safeguards."

Chief AI architect and Google DeepMind SVP Koray Kavukcuoglu said Argon delivers "frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense." Google is also "actively engaged in the U.S. government's voluntary process for pre-release model access," Kavukcuoglu added, per The Verge.

The announcement closes a long gap in Google's flagship lineup. The company hasn't released a new top-tier model since the Gemini 3 series in November 2025. Since then, OpenAI's GPT-6 and Anthropic's Mythos models have shipped and, by most third-party evals, pulled ahead. Google had promised a Gemini 3.5 Pro update in June but never shipped it, instead spending the summer releasing smaller Flash models — a move Kavukcuoglu described as taking "a little bit of a step back," according to The Verge's earlier reporting. Kavukcuoglu, who took over DeepMind after Demis Hassabis stepped down in August, told The Information in late September that Gemini 4 was in its refinement stage and that Google intended to release an early post-training output "as soon as possible" — much earlier than the end of the year.

For engineers, the most concrete spec so far is context. Ars Technica reports that Gemini 4 Argon supports a 1 million-token output limit, up from 64,000 tokens in previous Gemini models. That's a significant jump for long-horizon agentic tasks like analyzing entire codebases or generating large migrations in a single pass. Google hasn't announced API pricing yet, so cost-per-million-token remains unknown.

Google also shared early internal deployment numbers. According to Ars Technica, 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/C++ codebases to Rust, including thousands of lines in the core re2 and libgav1 libraries and more than 800,000 lines in the Fuchsia OS Zircon kernel. Those are meaningful proof points for teams considering agentic coding tools, though they come from Google's own infrastructure and should be read as such.

The benchmark chart Google released is characteristically aggressive. On the software engineering DeepSWE v1.1 benchmark, Argon scores 77.9 percent, above GPT-6 Astra, Fable 5.1, and Opus 5.5. Google also points to an industry-leading score on the economic analysis Vals Index. But these are vendor-reported numbers with no independent verification yet — the AI community spent the week before the announcement buzzing over what appeared to be leaked benchmarks on X.

Before a broader rollout, Google says it will strengthen defenses against misuse and prompt injection attacks and monitor for misalignment. The launch lands a day after OpenAI's DevDay, where the company introduced GPT-6.1 Sol and its Dots agent — and also shelved a planned GPT-6.1 Astra model over safety concerns.

For developers, the takeaway is straightforward: Gemini 4 Argon is real, but the public API is still a ways off. The 1 million-token output limit and the internal migration numbers suggest Google is aiming squarely at agentic coding and enterprise knowledge work. Until access broadens and independent evals land, treat the benchmark chart as a claim, not a verdict.