Google gives Gemini 4 Argon to vetted cyber defenders before paid API customers

Google gives the price as a ratio: half of Claude Opus 5.5, labelled introductory ^[4]. The dev.to write-up of the launch does not include per-token rates. Right now the price is the only part of Argon that everyone can use. A team can put the ratio into a cost model, but it cannot sign a contract on it while paid API access has no date ^[5].

The benchmark table needs the same caution. Google reported the scores itself, and the rival figures came from provider and public-leaderboard sources, not from identical test setups ^[10]. On DeepSWE v1.1, Argon's 77.9% is 3.7 points above Opus 5.5 and 3.8 points above GPT-6 Astra ^[11][17]. A lead that small carries over to another team's work only if all three models ran on the same harness, with comparable thinking budgets, on tasks like that team's backlog. According to the write-up, they did not share a setup ^[10].

Artificial Analysis offers a partial outside check. Its Intelligence Index puts Argon at 53, level with GPT-6 Astra at max settings ^[13]. Its Terminal-Bench 4 run has Argon at 57%, behind Sonnet 5.5 at 64%, Opus 5.5 at 60% and Astra at 59% ^[12]. On that agentic terminal work, Argon trails Sonnet by 7 points ^[18]. VentureBeat's tally also leaves five of its 18 benchmarks where Argon did not finish first ^[15].

The widest gap Google claims is in legal work. It reports 19.6% on Harvey Legal Agent, against 5.4% for Astra and 3.8% for Opus 5.5 ^[2]. That is about five times the Opus score ^[16]. Legal teams have the most reason to try to reproduce that result. Only Google's early-access groups can call the model at all ^[5].

The engineering change I would test first is the output ceiling ^[8]. The write-up argues that an agent holding its whole plan and deliverable in one generation does not need to chunk, resume and re-ground itself ^[19]. That is a real simplification of how agent pipelines are built today. Google disclosed almost nothing about architecture or parameter count ^[14]. Whether quality holds across a million output tokens therefore has to be measured on long runs.

My context is a team shipping on a fixed schedule. There, Argon goes on the eval list and stays off the purchasing plan. Google's own staff already use it, thousands of them, for coding, research and writing ^[7]. Google lists paid API customers as "next" ^[5]. Once that has a date, I would first run long-horizon coding tasks and legal drafting tasks from our own backlog on Argon and on the model we already run in production.