Gemini 4 Argon

Gemini 4 Argon earns a panel seat as the Google voice: 82% [71%–92%] recall at rank 1 of 17, measurably above 6 ranked models, with 0 measurably above it. If your other seats are Anthropic and OpenAI models, it adds a third maker's view, which is the point of a second opinion.

Its cost is time. A review took 99s, while Claude Sonnet 5.5 finished in 11s. Don't put it where a developer is waiting on the first comment. Put it in a background round.

A review cost $0.0527, 1.3× what GPT-6.1 Sol cost, for a recall the benchmark can't separate from Sol's. Argon lists at $2.00 / $10.00 per 1M tokens and Sol at $2.00 / $10.00. In this run Argon also counted more input tokens and wrote more output per review, and that moves the measured bill.

It raised 1.1 [0.6–1.7] false alarms per clean PR, and 78% of its catches sit on the defect's own lines. One of the three judges is from Google, so its vote is dropped for this model.

Its misses include "crypto/tls: add LocalCertificate to ConnectionState". The bars by kind and language rest on a few fixtures each, so don't read a strength into them. The methodology explains the intervals.