Gemini 4 Argon ties two rivals on a leading AI index while some Google staff doubt its coding

Five points separate Argon from Claude Opus 5.5 on the index, and three from Claude Sonnet 5.5 ^[1]. Google's strongest benchmark claims are narrower. Argon shares the top score with GPT-6 Astra on CWE-bench, a test of finding and patching security vulnerabilities, and Google says it set a new high on a test of long-horizon engineering work ^[13].

The doubts inside Google concern what those numbers leave out. One employee with direct access singled out front-end design, the work that decides how an app or website looks and feels ^[5]. Edwin Chen, founder of the AI startup Surge AI, speaking about benchmarks generally, described how relying on them can lead a lab to produce code in a given programming language while paying less attention to whether the finished app is easy to use or well designed ^[10]. "An analogy would be, 'Oh yeah, my kid got a really good score on the SAT', but the SAT doesn't translate into real-world performance," Chen said. "It's an incredibly pernicious problem." ^[11] The incentive starts with buyers. Implicator, summarizing Bloomberg's reporting, says labs prioritize these scores because customers often judge models by benchmark rankings ^[12].

Google's evidence comes from its own systems. "We've seen strong performance up close as Googlers have put the model through its paces in recent weeks, with many relying on it for their hardest coding and research problems," said Tulsee Doshi, who leads Gemini products at DeepMind ^[8]. In the launch example, Argon agents replaced 32,000 lines of SIMD code in a Rust port of Google's libgav1 video decoder, and Google said the result ran 2.7 times faster than the earlier Rust version with identical output ^[9]. Google also said the model helped its engineers free more than 300 tebibytes of memory across its data centers without new hardware ^[14]. Koray Kavukcuoglu, who runs DeepMind day to day, told a conference audience last week: "In my mind, it's a certainty that we are always gonna be at the frontier." ^[18]

A skeptic of the Bloomberg account would note that the critics are anonymous ^[15]. Google's staff also disagree among themselves, some believing Anthropic and OpenAI are improving faster and others that Google has caught up ^[16]. The objection is fair, and it applies to Google's side as well. A decoder rewrite and a memory clean-up are performance jobs on code Google knows well and chose to publish ^[9][14]. Neither is front-end work, the area one employee named ^[5]. The reports cite no results on code from outside Google.

I'd separate this quarter's decision from next quarter's. This quarter most teams cannot run Argon at all ^[1], and Google's last promised frontier date passed with no model ^[17]. The choice open now is whether to build an evaluation set from a team's own tickets and repositories before access arrives. When API customers do get in, a team with that set can judge Argon on its own code. A team without one will be choosing on a 53 that some of Google's own staff say overstates its coding ^[4].