Google's Gemini 4 Argon Tops the Benchmarks — and Almost No One Can Use It

Google DeepMind on September 30 unveiled Gemini 4 Argon, its first flagship model since the Gemini 3 series shipped last November, and immediately did something unusual for a model its own benchmark data says is the best on the market: it mostly kept it locked up. Rather than a broad consumer and API rollout, Argon is launching through Google's Fairwind Program, a vetted-access track for cybersecurity defenders, alongside the U.S. government's voluntary pre-release testing process. Broader API and Google AI Ultra access is planned only after what Google calls further guardrail strengthening.

The caution sits oddly next to the scorecard. Across 18 disclosed third-party benchmarks compiled by VentureBeat, MarkTechPost and 9to5Google, Argon leads or ties on 12 to 13, outperforming OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 on knowledge-work and automation tasks: 77.9% on the DeepSWE v1.1 coding benchmark versus 74.2% for Opus and 74.1% for Astra, 51.3% on AutomationBench against 42.5% and 41.4%, and a striking 19.6% on Harvey's Legal Agent benchmark versus roughly 4-5% for its rivals. It is not a clean sweep: Astra still beats Argon by more than ten points on FrontierSWE v2 (65.5% to 55.0%), and Opus edges it on Terminal-Bench Science.

A deliberately short leash

The restraint traces to what Argon was actually built to do. Google trained the model for autonomous vulnerability discovery and remediation, and says it has already surfaced critical flaws in healthcare software that prior frontier models missed. That capability cuts both ways: a system skilled enough to find and patch security holes at scale is, by the same token, skilled enough to find ways in. Google reports a 0.7% attack success rate on the Gray Swan indirect prompt-injection benchmark, a figure the company is using to justify handing the model to defenders before anyone else.

"Very excited to announce Gemini 4 Argon, our new Frontier AI model, and a huge step forward across key capabilities. We're focused on rolling it out responsibly starting with government and trusted cyber defenders through our Fairwind Program today." — Demis Hassabis, CEO, Google DeepMind

Argon also pushes maximum output from 64,000 tokens to 1 million per response, a jump Google frames as letting the model "think deeply and generate hundreds of thousands of tokens in a single trajectory" rather than breaking long tasks into chained calls. Pricing is already set for when the gate opens wider — $2 per million input tokens and $10 per million output tokens at introductory rates, rising to $4 and $20 — which signals Google intends a real commercial launch, not just a security showcase, once it decides the model is safe to hand to the open market.

Coming after a reported slip from an original June target and amid investor pressure over Google's release cadence, Argon reads less like a victory lap and more like a hedge: proof the lab can retake the benchmark lead, paired with an unusually public admission that leading the benchmarks and being ready for the world are no longer the same milestone.