Google grades its own Gemini 4 Argon 3.7 points ahead of Claude on long-horizon coding

Google computed Argon's 77.9% on DeepSWE itself, according to Decrypt ^[1][3]. The 74.2% for Claude Opus 5.5 and the 74.1% for GPT-6 Astra came from a public leaderboard and company reports ^[2][3]. So the lead over Anthropic's model is 3.7 points ^[1], and two different graders produced the scores on either side of it. Google's own table also concedes ground: Argon leads 12 of 18 benchmarks, ties one and trails on five ^[5].

The cyber claim changes with the publisher. Decrypt's headline says Argon "tops all other AI models on cybersecurity" ^[8]. CNBC reports Google saying the model ties for first, level with GPT-6 Astra and Grok 4.7 on cybersecurity evaluation benchmarks ^[9]. Argon's clearest lead is on Gray Swan's indirect prompt injection test, where attacks succeeded 0.7% of the time within 15 tries, against 1.0% for both Claude Opus 5.5 and Claude Fable 5.1 ^[6]. The gap is 0.3 points, or 30% fewer successful attacks ^[3]. GPT-6 Astra was tricked 8.5% of the time ^[7].

Revenue comes last in the release order. More than 650 Fairwind partners, including governments and critical infrastructure operators, get the model first and "without cyber guardrails" ^[10]. Paid API customers and Google AI Ultra subscribers come after ^[12]. When the API opens, the introductory rate of $2 per million input tokens and $10 per million output tokens is half the standard $4 and $20, and Google has not said when the discount ends ^[13][4]. The output ceiling is now 1 million tokens per reply, up from 64,000, a 15.6-fold rise ^[14][6]. One maximum-length reply costs $10 in output at the introductory rate and $20 at list ^[5].

The saving Google can already count sits inside its own fleet. The company told CNBC that Argon has been used to optimise data-center memory, freeing hundreds of terabytes without buying more hardware ^[16]. Argon follows a July in which Google shipped smaller Flash models but skipped the promised Gemini 3.5 Pro; Alphabet shares fell about 4.4% ^[17].

"Starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible," Tulsee Doshi, Google's Gemini model product lead, told CNBC ^[18]. On the Wiz Penetration Test Benchmark, an internal Google test of writing working exploits against web-application flaws, Argon solved 70.9% on the first try ^[15]. Google's restricted Gemini 3.8 Flash Cyber solved 58.2%, so the gain is 12.7 points ^[15][7]. Google says it will scale safeguards in four areas, including misuse and prompt injection, before launching publicly ^[19].

I think the coding lead points the right way but is small enough that a neutral grader could erase it. The case against that caution is the 28.9-point gain over the 49% Gemini 3.6 Flash scored in July ^[4][2]. A jump that size is hard to put down to grading method. If the public leaderboard confirms something near 77.9%, the caution is wrong and Google holds the top score until the next release. Claude Opus 5.5 shipped a week before Argon, and GPT 6.1 Sol a day before ^[11]. If the leaderboard puts Argon below 74.2%, the coding lead is gone and the gain over Flash is what remains.