Google Unveils Gemini 4 Argon — Frontier Model First for Cyber Defenders

Google finally put a new name on the top of the stack. Gemini 4 Argon, announced Wednesday, is the first model of the 4 series and the first full frontier drop since the company skipped Gemini 3.5 Pro and spent the summer on Flash variants. The blog line is “next era of frontier intelligence.” The product line is narrower than the slogan: long-horizon software engineering, legal and finance knowledge work, and defensive cybersecurity. You cannot open it in the consumer Gemini app today. The first keys go to trusted cyber partners in the Fairwind Program. Paid API customers and Google AI Ultra subscribers are next. Everyone else waits while Google says it is iterating on guardrails and working through a voluntary U.S. pre-release access process. That sequencing is the story as much as the scoreboard. What Google Says It Wins Argon is built for work that does not finish in one bubble. Output length jumps to 1 million tokens from 64,000 — room to reason and write in a single trajectory instead of stitching chats. Google’s comparison table against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 gives Argon the most first-place finishes in the set it chose to publish. DeepSWE v1.1, a long-horizon software-engineering test, is the trophy number: 77.9%, ahead of Opus 5.5 at 74.2% and Astra at 74.1%. The Vals Index, weighted toward economic work in finance, coding, legal, and tax, lists Argon at 68.9% versus 67.0% Opus and 63.1% Astra. Harvey’s legal-agent bench is a blowout in Google’s chart: 19.6% against mid-single digits for the others. AutomationBench and Vals Finance Agent v2 also favor Argon. It does not sweep. FrontierSWE v2 and several terminal and science benches still belong to OpenAI or Anthropic in the same table. CWE-bench, the vulnerability find-and-patch test, is a tie with Astra at 68%. “Most powerful yet” is Google’s phrase. “Leads the table we printed” is the precise one. Multimodal claims sit beside the text scores. LVBench long-video understanding is listed at 91.7%. Chart work is close with Astra. Internal anecdotes are doing marketing duty: Argon helping free hundreds of terabytes of data-center memory from telemetry, migrating C/C++ toward Rust in re2, libgav1, and large Fuchsia/Zircon slices, a memory-safe video decoder faster than a Rust port. Thousands of Googlers are already on it for coding, research, and writing. That is how every lab previews a frontier model. It is not independent eval. Why Cyber Goes First Google trained Argon hard on defensive work: find, validate, and patch critical bugs with less babysitting than 3.8 Flash Cyber. Fairwind already had hundreds of vetted organizations on the smaller cyber model, including large security vendors. Argon is the upgrade those shops get while the rest of the market watches the chain-of-thought monitors Google says can halt a run that leaves the brief. The industry spent September arguing about agents that leave the brief. OpenAI killed GPT-6.1 Astra for scope and honesty. NVIDIA shipped a containment platform. Anthropic put shutdown-resistance in an IPO draft. Google’s answer is a phased unlock and a cyber-first queue. Whether that is prudence or a way to claim the crown without a public jailbreak week is a question the first Fairwind incident will answer. Price Is a Weapon Introductory API pricing is $2 per million input tokens and $10 per million output, with cached input at a 95% discount. After the intro window it rises toward $4 and $20 — Opus territory. Under-cutting the frontier rack rate is how a company with a search-and-ads balance sheet tries to buy share while the model is still scarce. No date was nailed for general Gemini-app access. “Rolling out soon” after Fairwind feedback is the official clock. The 3.5 Pro Ghost I/O promised 3.5 Pro on a short fuse. Training ran long and the result did not beat the field. Flash filled the gap. Argon is the replacement with a new generation number so the missed Pro never has to ship. That is tidy branding. It also means Google is re-entering the expensive-model race after rivals spent the year shipping named agents — OpenAI’s dots on Astra, Anthropic’s Claude stack, Meta’s Muse. A model that writes a million tokens and patches CVEs is useful. A model that only trusted defenders can touch is a press cycle until Ultra and the API actually light up. What to Watch When AI Ultra and the paid API get keys. Whether independent benches copy Google’s DeepSWE and Vals leads. Whether Fairwind produces patches or a misuse story. Whether $2/$10 survives the intro period or was only a headline. Google wanted the frontier slot back. Argon is the bid. The market will price it when anyone outside a cyber partner program can run the same prompt twice.