Google Launches Gemini 4 Argon to Cyber Defenders First as Benchmark Claims Face Staff Skepticism

Google on Wednesday introduced Gemini 4 Argon, its most capable artificial-intelligence model, and said it would roll out first to a small group of trusted cyber defenders - a deliberately restricted debut that landed one day after the company signed the White House's "morally binding" accord on AI self-policing.

A velvet-rope debut

Google said Argon is "rolling out to a set of trusted cyber defenders through our Fairwind Program," in an announcement written by DeepMind executive Koray Kavukcuoglu. The trusted cohort gets the model "without cyber guardrails," so defenders can use its full frontier-level capabilities. Wiz, the cloud-security firm, is already using Argon in its Scan for Good initiative; Google says the model uncovered a critical vulnerability in healthcare software used by hospitals worldwide, "a severe risk that previous frontier models had missed."

The company also says it is "actively engaged in the U.S. government's voluntary process for pre-release model access," and will expand access gradually - "starting with paid API customers and Google AI Ultra subscribers" - "as soon as possible." No dates appear anywhere in the post, and Reuters reports the company gave no timeline for a public release.

The specs Google is selling

The model's output token limit jumps to 1 million tokens, up from 64,000, which Google calls industry-leading. It launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced 95% below the input-token price. Artificial Analysis reports the standard price is $4 per million input tokens and $20 per million output tokens, discounted 50% "for at least one month" - Google has not confirmed when the promotion ends.

Google is also dogfooding Argon at scale. Its agents migrated the 800,000-plus-line Fuchsia OS Zircon kernel from C and C++ to Rust, rewrote a video decoder so that a memory-safe Rust version runs 2.7 times faster than the earlier Rust port with identical video output, and identified memory optimizations expected to free more than 300 TiB across its data centers once rolled out, with estimated total savings of 500 TiB to 1 PiB. A tebibyte is roughly 1.1 terabytes. For researchers, Google adds one more data point: Argon beat a published baseline by 40% "in a matter of minutes" on a spacetime optimization problem for quantum subroutines.

The benchmarks, and the skeptics

On Artificial Analysis's Intelligence Index, Argon scores 53 at its highest reasoning setting - level with OpenAI's GPT-6 Astra at maximum and Anthropic's Claude Fable 5.1, and behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56. The outfit's verdict: "Google is back as one of the top three labs in intelligence achieved." It also reports a 15% hallucination rate for Argon, the lowest of any model scoring 45 or higher, against 51% for GPT-6 Astra at max and 54% for GPT-6.1 Sol. On Terminal Bench 4, a coding benchmark, Argon hits 57%, behind Claude Sonnet 5.5 at 64% and Opus 5.5 at 60%.

Not everyone inside Google is buying the numbers. Bloomberg reported that some employees with direct access say the model does less well on real work, coding in particular, than its benchmark scores suggest - a practice critics call "benchmaxxing," concentrating engineering effort on test scores rather than on whether the product does the user's job well. One insider singled out front-end design as a weak spot, and two people told Bloomberg the model appears affected by benchmaxxing. Some employees believed Anthropic's Fable and OpenAI's Astra models were improving faster than Gemini, and that Gemini 4, even at its best, would still lag behind those models in some areas; others thought Google had caught up. Google said it would be "inaccurate" to say the model underperforms in coding, and one employee cited "large consensus" internally that Argon sits at the frontier.

The accord's first live test

Tuesday's White House accord committed Google and five other AI leaders to a four-layer safety regime: internal controls to monitor model capabilities and alignment, watching for cybersecurity, biosecurity and chemical threats; a dedicated internal team to make sure those controls work; an independent external auditor to verify them; and a board committee to receive the auditors' reports and ensure problems get fixed. The enforcement mechanism is the one President Trump named: moral obligation, backed by public pressure. There are no penalties, no deadlines, and no audit schedule; the text says only that it "may make sense" to codify the steps into law over time. The accord was signed in the East Room on Tuesday by Google chief executive Sundar Pichai, Anthropic's Dario Amodei, Meta's Mark Zuckerberg, OpenAI president Greg Brockman, Nvidia's Jensen Huang and xAI founder Elon Musk.

TickerGrove's read: Argon is the first live test of that posture. The accord text never mentions staged rollouts or pre-release government access - Google's restricted debut is a company choice, not an accord clause. But the sequence is the point: one day after Pichai signed a pledge to police frontier risk, Google chose to ship its most capable model behind a defender-first velvet rope, with Washington in the loop before the public. If the accord is going to mean anything before its first audit is ever scheduled, this is what it looks like.

What decides the next chapter

Whether a public release date ever materializes, and what happens to pricing when the introductory discount expires - on both, Google has given no end date. Whether real-world coding performance converts the internal skeptics, or the benchmaxxing question follows the model into enterprise deployments. And the backdrop: Google has scrapped the planned Gemini 3.5 Pro, which Pichai had said would arrive in June, while co-founder Demis Hassabis stepped aside as DeepMind's chief executive in August to become chair.