Google released Gemini 4 Argon Wednesday, but only to a vetted group of cyber defenders, according to a post on the Google Blog by Koray Kavukcuoglu, senior vice president of Google DeepMind and the company's chief AI architect. Google pitches the model at long-running software engineering, legal and financial work and vulnerability hunting, and says it outscores rival frontier systems on most of the tests it published.

What Happened

Argon is going first to participants in Fairwind, a Google security program for trusted defenders. Google wrote that for those defenders and its own internal teams it is releasing Argon without cyber guardrails, so they can use its full capability to find and fix flaws. Google says the model can locate, confirm and patch critical software vulnerabilities on its own. As an example, it said security firm Wiz used Argon through its Scan for Good initiative to uncover a critical flaw in healthcare software used by hospitals worldwide that exposed sensitive personal data, an issue earlier frontier models had missed.

The model's maximum output grows to 1 million tokens, from a previous ceiling of 64,000. Google-reported results include 77.9% on DeepSWE v1.1, 68% on the CWE-bench v1 vulnerability-remediation test, 91.7% on the LVBench long-video benchmark and a first-place 51.3% on AutomationBench. TechCrunch noted that benchmarking startup Vals currently ranks Argon first on its model index.

RelatedAnthropic Prices Claude Haiku 5.5 at a Tenth of Its Predecessor for Most API Requests

Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. Those rates later rise to $4 and $20.

Why It Matters

The rollout order puts security use ahead of general availability, and the decision to hand some users an unrestricted version is unusual for a frontier launch. Google listed safeguards it says will apply more broadly: misuse refusals backed by monitoring of the model's internal activations, prompt-injection hardening, monitoring of chain-of-thought for misalignment and sandboxed testing of high-risk capabilities. VentureBeat reported that Argon is also going through the U.S. government's voluntary pre-release model access process.

Independent comparisons give a more mixed picture than the headline claims. VentureBeat counted 18 disclosed benchmarks and found Argon first on 12 and tied on one. It trails GPT-6 Astra by 10.5 points on FrontierSWE v2 (55.0% against 65.5%) and on Terminal-Bench Science (57.6% against 68.1%). On Harvey's legal-agent benchmark it posted 19.6%, against 5.4% for Astra and 3.8% for Anthropic's Claude Opus 5.5. In Gray Swan's indirect prompt-injection test, attacks succeeded 0.7% of the time against Argon, compared with 1.0% for Anthropic's models and 8.5% for Astra.

Price is the other lever. VentureBeat calculated that the launch rates are one-fifth of Astra's listed $10 and $50, and that the later rates match Claude Opus 5.5.

How Google Uses It Internally

Google described several internal projects. Argon has helped migrate C and C++ code to Rust, including more than 800,000 lines for the Zircon kernel in Fuchsia. A memory-efficiency effort freed more than 300 TiB across Google's data centers. A SIMD rewrite of the libgav1 video decoder, guided by profiling experiments, ran 2.7 times faster.

Context

Google DeepMind said last week that Gemini 4 was nearing release; Argon is the first model in the family to ship. It follows OpenAI's GPT-6 Astra, released in early September, and Anthropic's Fable, released earlier this year. TechCrunch noted that the Gemini app passed 1 billion monthly users in August.

What to Watch

Google said broader access will begin with paid API customers and Google AI Ultra subscribers, but did not say when. Outside testing will show whether its benchmark lead holds up, and security researchers will be watching how Google controls the guardrail-free version given to Fairwind participants.