What Happened
Google has finally put a new flagship model on the table. On September 30, the company announced Gemini 4 Argon, its first frontier-class model since Gemini 3 arrived in November 2025, and it did something unusual with the release: it handed the model to a small group of cybersecurity professionals first. Access is currently limited to vetted defenders in Google's Fairwind Program, according to The Next Web. Paid API customers and Google AI Ultra subscribers are next in line, with no public date attached.
The headline numbers are strong in the places Google chose to emphasize. Argon scores 77.9% on DeepSWE v1.1, a software engineering benchmark, and 68% on CWE-bench, which measures how well a model fixes security flaws, a result that ties for first place. It also raises the output ceiling to 1 million tokens per response, up from the 64,000-token limit of earlier Gemini models. Google describes the model as built to sustain deep reasoning across long, complex workflows in areas like software engineering, legal work, finance and security.
Pricing is aggressive on paper. Introductory rates are $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 once the promotional period ends, based on a launch breakdown from ExplainX. Cached input gets a steep discount, and no end date for the introductory pricing has been announced. There is one important catch. As of October 1, Argon had not appeared in Google's public API catalog, so developers can read about the model but cannot yet build on it.
The wait has been long. Google confirmed on July 21 that it had started its most ambitious pre-training run yet. Until Argon reaches ordinary accounts, Gemini 3.8 Flash remains the newest model most people can actually use. Artificial Analysis scores Argon at 53 on its Intelligence Index, compared with 41 for that Flash model, so the jump from what's available today is large.
Why It Matters
The release matters less for any single benchmark than for what it says about Google's position. For most of the past year, the frontier conversation has been dominated by OpenAI and Anthropic, and Google has been the company with the distribution, the chips and the research talent but without the model everyone wanted to use. Argon is the first serious attempt to change that story.
The staged rollout is a strategic choice as much as a safety one. Putting a model with relaxed guardrails in front of vetted security teams lets Google test its cyber capabilities where the benefits are clearest and the audience is most accountable. Wiz, the cloud security firm, reportedly used Argon to find a critical healthcare vulnerability that earlier frontier models had missed. That is the kind of concrete result that makes a restricted launch look deliberate rather than delayed.
Coding is the commercial prize. R&D World cites Menlo Ventures estimates that Anthropic holds 54% of enterprise spending on coding models, while Google holds 11%. Coding agents are among the stickiest AI products companies buy, so closing even part of that gap would matter far more to Google's cloud business than a leaderboard win. The 1-million-token output limit is aimed squarely at that market, because long autonomous coding sessions are where short output windows hurt the most.
Reaction
The reception has been respectful but divided. On the plus side, the model took the top spot on Arena.ai's text leaderboard on October 1 with a score of 1525, according to the ExplainX roundup, and it leads on several long-context and agentic measures, including 84.2% on GraphWalks at 256K to 1M tokens. It also looks cost-efficient in practice. Artificial Analysis calculated that Argon costs $1.99 per reasoning task at launch pricing, versus $3.26 for OpenAI's GPT-6 Astra, even though Argon burns more output tokens to get there.
The caveats are just as visible. R&D World reports that Argon roughly matches GPT-6 Astra on Artificial Analysis's overall index, sits behind Claude Opus 5.5, and ranks eighth on Arena's Agent Arena leaderboard. Science results are split. On LABBench2, which tests literature research and experimental troubleshooting, Argon scores 88.8% against 85.4% for Astra and 73.1% for Opus 5.5. On Terminal-Bench Science, which requires completing computational research workflows, it trails at 57.6% against 68.1% and 63.3%. Strong at reading and reasoning, weaker at executing long tasks end to end.
Inside Google, the picture is mixed too. Bloomberg reported that some employees questioned how well Argon handles practical coding tasks, while others believed it had caught up with the leading rivals. Company leadership disputes the underperformance claims. Wall Street's mood, meanwhile, has shifted. CNBC's coverage framed the launch against investors who now want a breakout personal agent more than another model, a reminder that raw intelligence is no longer the whole pitch.
What's Next
The next milestone is access. Paid API customers and Google AI Ultra subscribers are the stated next stage, and the real test begins when ordinary developers can run Argon against their own codebases. Benchmarks like DeepSWE reward clean, well-specified problems. Production repositories are messier, and the employee skepticism Bloomberg described will either fade or harden once outside teams publish their own results.
Compute will shape the pace. A model that can emit a million tokens in a single response, at a launch price this low, puts real strain on infrastructure, and Google has said little about when capacity will be ready for general release. That is also where Google's advantages lie: it designs its own chips and runs its own data centers, which gives it room to hold prices down in a way that rivals renting capacity cannot easily match. The introductory pricing will be worth watching. If the $2 and $10 rates survive, they pressure everyone else's margins. If they jump to $4 and $20 quickly, the bargain was mostly a marketing window.
Expect the competition to answer fast. OpenAI's GPT-6 line and Anthropic's Claude models both lead Argon on some coding and terminal benchmarks, and neither company tends to leave a gap open for long. The more interesting question is whether Google can turn a respectable model into a product people reach for daily, since that is the gap analysts keep pointing at.
Closing Thoughts
Argon is a credible comeback, not a knockout. It is the best Gemini yet, it leads in some of the areas that matter, and the cyber-first rollout shows a company thinking carefully about who gets a powerful tool first. But it arrives into a market where the leaders have had months to entrench, and where customers have learned to distrust launch-day charts.
The most useful way to read this launch is as a bet on patience. Google took its time, shipped to specialists first, and priced aggressively to win trial. Whether that earns back the developers it lost depends on a boring factor: how the model behaves on a Tuesday afternoon, inside a real codebase, when nobody is watching the leaderboard. That answer is still a few weeks away.
Sources: The Next Web, R&D World, ExplainX, CNBC
한글 요약
구글이 9월 30일 제미나이 3 이후 처음으로 프런티어급 모델 「제미나이 4 아르곤(Gemini 4 Argon)」을 공개했습니다. 현재는 검증된 사이버 방어 전문가가 참여하는 페어윈드 프로그램에서만 쓸 수 있고, 유료 API 고객과 구글 AI 울트라 구독자가 그다음 순서입니다. DeepSWE v1.1에서 77.9%, 보안 결함 수정 벤치마크 CWE-bench에서 68%(공동 1위)를 기록했고, 출력 한도는 기존 6만 4천 토큰에서 100만 토큰으로 늘었습니다. 도입 가격은 입력 100만 토큰당 2달러, 출력 10달러이며 이후 4달러와 20달러로 오릅니다.
이번 출시는 구글의 입지를 보여 줍니다. 멘로 벤처스는 기업의 코딩 모델 지출에서 앤트로픽이 54%, 구글이 11%를 차지한다고 추정했습니다. 코딩 시장의 격차를 좁히는 것이 구글에는 가장 큰 상업적 과제이고, 100만 토큰 출력은 긴 자율 코딩 작업을 겨냥한 설계입니다. 보안 업체 위즈가 아르곤으로 이전 모델이 놓친 의료 분야 치명적 취약점을 찾았다는 보도도 제한 출시가 신중한 선택임을 뒷받침합니다.
평가는 엇갈립니다. 아레나닷에이아이 텍스트 순위 1위(1,525점)에 올랐고 추론 1건당 비용도 GPT-6 아스트라보다 낮다는 분석이 있지만, 종합 지수는 아스트라와 비슷하고 클로드 오퍼스 5.5에는 못 미치며 에이전트 아레나는 8위입니다. 블룸버그는 일부 구글 직원이 실제 코딩 성능에 의문을 제기했다고 전했고 경영진은 이를 부인했습니다. 앞으로는 일반 개발자에게 열리는 시점과 도입 가격 유지 여부가 관건이며, 실제 코드베이스에서의 성능이 이번 출시의 성패를 가를 전망입니다.